diff --git a/common/CMakeLists.txt b/common/CMakeLists.txt
index 54691da3f41e..36f1e0cd50f1 100644
--- a/common/CMakeLists.txt
+++ b/common/CMakeLists.txt
@@ -81,6 +81,8 @@ add_library(${TARGET}
     imatrix-loader.cpp
     imatrix-loader.h
     json-schema-to-grammar.cpp
+    json.cpp
+    json.h
     llguidance.cpp
     log.cpp
     log.h
diff --git a/common/arg.cpp b/common/arg.cpp
index 3da1d61f4278..86f8610a56d0 100644
--- a/common/arg.cpp
+++ b/common/arg.cpp
@@ -5,6 +5,7 @@
 #include "common.h"
 #include "download.h"
 #include "json-schema-to-grammar.h"
+#include "json.h"
 #include "llama.h"
 #include "log.h"
 #include "sampling.h"
@@ -21,9 +22,6 @@
 #include <shellapi.h>
 #endif
 
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
-
 #include <algorithm>
 #include <cinttypes>
 #include <climits>
@@ -32,6 +30,7 @@
 #include <filesystem>
 #include <fstream>
 #include <list>
+#include <numeric>
 #include <regex>
 #include <set>
 #include <string>
@@ -55,7 +54,7 @@
 
 #define LLAMA_MAX_URL_LENGTH 2084 // Maximum URL Length in Chrome: 2083
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 using namespace common_arg_utils;
 
 static std::initializer_list<enum llama_example> mmproj_examples = {
diff --git a/common/chat-auto-parser-generator.cpp b/common/chat-auto-parser-generator.cpp
index af84ff323daf..d7e117e4d98b 100644
--- a/common/chat-auto-parser-generator.cpp
+++ b/common/chat-auto-parser-generator.cpp
@@ -5,13 +5,12 @@
 #include "common.h"
 #include "json-schema-to-grammar.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <stdexcept>
 #include <string>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 // Helper to iterate over tools/functions
 static void foreach_function(const json & tools, const std::function<void(const json &)> & fn) {
@@ -391,7 +390,7 @@ common_peg_parser analyze_tools::build_tool_parser_tag_tagged(parser_build_conte
 
         std::set<std::string> required;
         if (params.contains("required")) {
-            params.at("required").get_to(required);
+            required = params.at("required").get<std::set<std::string>>();
         }
 
         auto schema_info = common_schema_info();
diff --git a/common/chat-auto-parser-helpers.cpp b/common/chat-auto-parser-helpers.cpp
index 81b17e5e1d27..b37906bdf85a 100644
--- a/common/chat-auto-parser-helpers.cpp
+++ b/common/chat-auto-parser-helpers.cpp
@@ -4,14 +4,11 @@
 #include "chat-peg-parser.h"
 #include "chat.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <cctype>
 #include <numeric>
 
-using json = nlohmann::ordered_json;
-
 std::string trim_whitespace(const std::string & str) {
     size_t start = 0;
     while (start < str.length() && std::isspace(static_cast<unsigned char>(str[start]))) {
diff --git a/common/chat-auto-parser.h b/common/chat-auto-parser.h
index 074216b11ee1..8ae15c91e1e5 100644
--- a/common/chat-auto-parser.h
+++ b/common/chat-auto-parser.h
@@ -4,7 +4,7 @@
 #include "common.h"
 #include "jinja/caps.h"
 #include "peg-parser.h"
-#include "nlohmann/json.hpp"
+#include "json.h"
 
 #include <chrono>
 #include <optional>
@@ -12,7 +12,7 @@
 #include <utility>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 class common_chat_peg_builder;
 
diff --git a/common/chat-diff-analyzer.cpp b/common/chat-diff-analyzer.cpp
index d6d2af2d50fe..a7e370578fd9 100644
--- a/common/chat-diff-analyzer.cpp
+++ b/common/chat-diff-analyzer.cpp
@@ -4,11 +4,11 @@
 #include "chat.h"
 #include "common.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
 #include "peg-parser.h"
 
 #include <algorithm>
 #include <cctype>
+#include <numeric>
 #include <ostream>
 #include <sstream>
 
@@ -17,7 +17,7 @@
 #define ANSI_ORANGE "\033[1m\x1b[38;5;214m"
 #define ANSI_RED    "\033[1m\x1b[38;5;196m"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 namespace autoparser {
 
@@ -929,7 +929,7 @@ void analyze_tools::analyze_tool_call_format_json_native(const std::string & cle
     int  json_end       = clean_haystack.find_last_of('}');
     std::string cut     = clean_haystack.substr(json_start, json_end - json_start + 1);
     json call_struct    = json::parse(cut);
-    auto register_field = [&](const std::string & prefix, const nlohmann::detail::iteration_proxy_value<json::iterator> & subel) {
+    auto register_field = [&](const std::string & prefix, const common_json_entry & subel) {
         if (subel.value().is_string() && std::string(subel.value()).find("call0000") != std::string::npos) {
             format.id_field = !prefix.empty() ? prefix + "." + subel.key() : subel.key();
         } else if (subel.value().is_string() && std::string(subel.value()) == fun_name_needle) {
diff --git a/common/chat-peg-parser.cpp b/common/chat-peg-parser.cpp
index 06737b165c05..79b97a80f1b2 100644
--- a/common/chat-peg-parser.cpp
+++ b/common/chat-peg-parser.cpp
@@ -4,12 +4,10 @@
 #include "ggml.h"
 #include "peg-parser.h"
 
-#include <nlohmann/json.hpp>
-
 #include <cstdint>
 #include <functional>
 
-using ordered_json = nlohmann::ordered_json;
+using ordered_json = common_json;
 
 static std::string_view trim_trailing_space(std::string_view sv, int max = -1) {
     int count = 0;
diff --git a/common/chat-peg-parser.h b/common/chat-peg-parser.h
index 5d764dbaa0ec..114fa049fa74 100644
--- a/common/chat-peg-parser.h
+++ b/common/chat-peg-parser.h
@@ -128,7 +128,7 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // parameters_order: order in which JSON fields should be parsed
     common_peg_parser standard_json_tools(const std::string &              section_start,
                                           const std::string &              section_end,
-                                          const nlohmann::ordered_json &   tools,
+                                          const common_json &   tools,
                                           bool                             parallel_tool_calls,
                                           bool                             force_tool_calls,
                                           const std::string &              name_key = "",
@@ -143,13 +143,13 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     // Legacy-compatible helper for building XML/tagged style tool calls
     // Used by tests and manual parsers
     common_peg_parser standard_constructed_tools(const std::map<std::string, std::string> & markers,
-                                                 const nlohmann::ordered_json &             tools,
+                                                 const common_json &             tools,
                                                  bool                                       parallel_tool_calls,
                                                  bool                                       force_tool_calls);
 
     // Helper for Python-style function call format: name(arg1="value1", arg2=123)
     // Used by LFM2 and similar templates
-    common_peg_parser python_style_tool_calls(const nlohmann::ordered_json & tools,
+    common_peg_parser python_style_tool_calls(const common_json & tools,
                                               bool                           parallel_tool_calls,
                                               bool                           allow_json_literals);
 
@@ -158,19 +158,19 @@ class common_chat_peg_builder : public common_peg_parser_builder {
     common_peg_parser python_or_json_value();
 
     // Implementation helpers for standard_json_tools — one per JSON tool call layout mode
-    common_peg_parser build_json_tools_function_is_key(const nlohmann::ordered_json & tools,
+    common_peg_parser build_json_tools_function_is_key(const common_json & tools,
                                                        const std::string &            args_key,
                                                        const std::string &            effective_args_key,
                                                        const std::string &            call_id_key,
                                                        const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_nested_keys(const nlohmann::ordered_json & tools,
+    common_peg_parser build_json_tools_nested_keys(const common_json & tools,
                                                    const std::string &            effective_name_key,
                                                    const std::string &            effective_args_key,
                                                    const std::string &            call_id_key,
                                                    const std::string &            gen_call_id_key);
 
-    common_peg_parser build_json_tools_flat_keys(const nlohmann::ordered_json &   tools,
+    common_peg_parser build_json_tools_flat_keys(const common_json &   tools,
                                                  const std::string &              effective_name_key,
                                                  const std::string &              effective_args_key,
                                                  const std::string &              call_id_key,
diff --git a/common/chat.cpp b/common/chat.cpp
index 39761f12acdd..24618d35aee5 100644
--- a/common/chat.cpp
+++ b/common/chat.cpp
@@ -6,6 +6,7 @@
 #include "common.h"
 #include "ggml.h"
 #include "json-schema-to-grammar.h"
+#include "json.h"
 #include "log.h"
 
 #include "jinja/value.h"
@@ -13,14 +14,13 @@
 #include "jinja/caps.h"
 #include "peg-parser.h"
 
-#include "nlohmann/json.hpp"
-
 #include <algorithm>
 #include <cstdio>
 #include <cstdlib>
 #include <ctime>
 #include <exception>
 #include <functional>
+#include <iomanip>
 #include <map>
 
 #include <optional>
@@ -30,7 +30,7 @@
 #include <utility>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static std::string format_time(const std::chrono::system_clock::time_point & now, const std::string & format) {
     auto               time       = std::chrono::system_clock::to_time_t(now);
@@ -48,7 +48,7 @@ static json safe_args_parse(const std::string & to_parse) {
     }
     try {
         return json::parse(stripped);
-    } catch (json::exception & e) {
+    } catch (const common_json_error & e) {
         return stripped;
     }
 }
@@ -488,17 +488,17 @@ struct messages_inp_normalizer {
         json normalized = json::array();
         for (const auto & msg : messages) {
             json copy = msg;
-            auto it = copy.find("content");
-            if (it != copy.end()) {
-                if (only_typed && it->is_string()) {
-                    *it = json::array({
+            if (copy.contains("content")) {
+                json & it = copy.at("content");
+                if (only_typed && it.is_string()) {
+                    it = json::array({
                         json{
                             {"type", "text"},
-                            {"text", it->get<std::string>()},
+                            {"text", it.get<std::string>()},
                         }
                     });
-                } else if (only_string && it->is_array()) {
-                    *it = concat_content_parts(*it);
+                } else if (only_string && it.is_array()) {
+                    it = concat_content_parts(it);
                 }
             }
             normalized.push_back(std::move(copy));
@@ -608,7 +608,7 @@ std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const json & too
     return result;
 }
 
-common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value) {
+common_chat_continuation common_chat_continuation_parse(const common_json & value) {
     if (value.is_boolean() && value.get<bool>()) {
         return COMMON_CHAT_CONTINUATION_AUTO;
     }
@@ -920,7 +920,7 @@ static void foreach_parameter(const json &
     const auto &          props = params.at("properties");
     std::set<std::string> required;
     if (params.contains("required") && params.at("required").is_array()) {
-        params.at("required").get_to(required);
+        required = params.at("required").get<std::set<std::string>>();
     }
     for (const auto & [name, prop] : props.items()) {
         bool is_required = (required.find(name) != required.end());
@@ -937,7 +937,7 @@ static std::string common_chat_template_direct_apply_impl(
     jinja::context ctx(tmpl.source());
 
     // messages_override is already built for this template, do not touch its content parts
-    nlohmann::ordered_json inp = nlohmann::ordered_json{
+    json inp = json{
         {"messages", messages_override.has_value()
             ? *messages_override
             : messages_inp_normalizer(tmpl.original_caps()).normalize(inputs.messages)},
@@ -1058,7 +1058,7 @@ static common_chat_params common_chat_params_init_ministral_3(const common_chat_
                 });
             } else if (msg.at("content").is_array()) {
                 auto blocks = msg.at("content");
-                content.insert(content.end(), blocks.begin(), blocks.end());
+                content.insert(blocks);
             }
         }
 
@@ -2238,7 +2238,7 @@ static common_chat_params common_chat_params_init_deepseek_v3_2(const common_cha
 
                 std::set<std::string> required;
                 if (params.contains("required")) {
-                    params.at("required").get_to(required);
+                    required = params.at("required").get<std::set<std::string>>();
                 }
 
                 auto schema_info = common_schema_info();
@@ -2860,7 +2860,7 @@ static common_chat_params common_chat_params_init_minimax_m3(const common_chat_t
 
                 std::set<std::string> required;
                 if (schema.contains("required")) {
-                    schema.at("required").get_to(required);
+                    required = schema.at("required").get<std::set<std::string>>();
                 }
 
                 std::vector<common_peg_parser> required_elements;
@@ -2972,10 +2972,10 @@ static void system_message_not_supported(json & messages) {
             auto & second_msg = messages[1];
             second_msg["content"] = first_msg.at("content").get<std::string>()
                 + "\n" + second_msg.at("content").get<std::string>();
-            messages.erase(messages.begin());
+            messages.erase(0);
         } else {
             LOG_WRN("Removing system prompt due to template not supporting system role\n");
-            messages.erase(messages.begin());
+            messages.erase(0);
         }
     }
 }
diff --git a/common/chat.h b/common/chat.h
index 6d5b220aebb5..cb39e3458f44 100644
--- a/common/chat.h
+++ b/common/chat.h
@@ -8,7 +8,7 @@
 #include "jinja/runtime.h"
 #include "jinja/caps.h"
 
-#include "nlohmann/json_fwd.hpp"
+#include "json.h"
 
 #include <chrono>
 #include <functional>
@@ -17,7 +17,6 @@
 #include <vector>
 
 using chat_template_caps = jinja::caps;
-using json = nlohmann::ordered_json;
 
 struct common_chat_templates;
 
@@ -87,7 +86,7 @@ struct common_chat_msg {
     std::string                               tool_name;
     std::string                               tool_call_id;
 
-    nlohmann::ordered_json to_json_oaicompat(bool concat_typed_text = false) const;
+    common_json to_json_oaicompat(bool concat_typed_text = false) const;
 
     std::string render_content(const std::string & delimiter = "\n\n") const;
 
@@ -211,7 +210,7 @@ struct common_chat_msg_delimiters {
     // split tokens into message spans. skips maps a start index to a length of a region to jump over without matching
     common_chat_msg_spans split(const llama_tokens & tokens, const std::map<size_t, size_t> & skips = {}) const;
 
-    nlohmann::ordered_json to_json() const;
+    common_json to_json() const;
 };
 
 struct common_chat_tool {
@@ -350,16 +349,16 @@ common_chat_tool_choice common_chat_tool_choice_parse_oaicompat(const std::strin
 bool common_chat_templates_support_enable_thinking(const common_chat_templates * chat_templates);
 
 // Parses a JSON array of messages in OpenAI's chat completion API format.
-std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const nlohmann::ordered_json & messages);
+std::vector<common_chat_msg> common_chat_msgs_parse_oaicompat(const common_json & messages);
 
-std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const nlohmann::ordered_json & tools);
+std::vector<common_chat_tool> common_chat_tools_parse_oaicompat(const common_json & tools);
 
-common_chat_continuation common_chat_continuation_parse(const nlohmann::ordered_json & value);
+common_chat_continuation common_chat_continuation_parse(const common_json & value);
 
 // DEPRECATED: only used in tests
-nlohmann::ordered_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
+common_json common_chat_msgs_to_json_oaicompat(const std::vector<common_chat_msg> & msgs, bool concat_typed_text = false);
 
-nlohmann::ordered_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
+common_json common_chat_tools_to_json_oaicompat(const std::vector<common_chat_tool> & tools);
 
 // get template caps, useful for reporting to server /props endpoint
 std::map<std::string, bool> common_chat_templates_get_caps(const common_chat_templates * chat_templates);
@@ -386,4 +385,4 @@ struct common_chat_prompt_preset {
 
 common_chat_prompt_preset common_chat_get_asr_prompt(const common_chat_templates * chat_templates);
 
-common_chat_msg_delimiters common_chat_msg_delimiters_parse(const nlohmann::ordered_json & delimiters);
+common_chat_msg_delimiters common_chat_msg_delimiters_parse(const common_json & delimiters);
diff --git a/common/download.cpp b/common/download.cpp
index 2509f75ab996..4b28a708c86e 100644
--- a/common/download.cpp
+++ b/common/download.cpp
@@ -5,9 +5,7 @@
 #include "log.h"
 #include "download.h"
 #include "hf-cache.h"
-
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <algorithm>
 #include <filesystem>
@@ -44,8 +42,6 @@
 #include <unistd.h>
 #endif
 
-using json = nlohmann::ordered_json;
-
 //
 // downloader
 //
@@ -856,8 +852,8 @@ static std::string common_docker_get_token(const std::string & repo) {
         throw std::runtime_error("Failed to get Docker registry token, HTTP code: " + std::to_string(res.first));
     }
 
-    std::string            response_str(res.second.begin(), res.second.end());
-    nlohmann::ordered_json response = nlohmann::ordered_json::parse(response_str);
+    std::string response_str(res.second.begin(), res.second.end());
+    common_json response = common_json::parse(response_str);
 
     if (!response.contains("token")) {
         throw std::runtime_error("Docker registry token response missing 'token' field");
@@ -919,9 +915,9 @@ std::string common_docker_resolve_model(const std::string & docker) {
             throw std::runtime_error("Failed to get Docker manifest, HTTP code: " + std::to_string(manifest_res.first));
         }
 
-        std::string            manifest_str(manifest_res.second.begin(), manifest_res.second.end());
-        nlohmann::ordered_json manifest = nlohmann::ordered_json::parse(manifest_str);
-        std::string            gguf_digest;  // Find the GGUF layer
+        std::string manifest_str(manifest_res.second.begin(), manifest_res.second.end());
+        common_json manifest = common_json::parse(manifest_str);
+        std::string gguf_digest;  // Find the GGUF layer
         if (manifest.contains("layers")) {
             for (const auto & layer : manifest["layers"]) {
                 if (layer.contains("mediaType")) {
diff --git a/common/hf-cache.cpp b/common/hf-cache.cpp
index f1dacaa4778d..50d6dd6105c4 100644
--- a/common/hf-cache.cpp
+++ b/common/hf-cache.cpp
@@ -4,9 +4,7 @@
 #include "common.h"
 #include "log.h"
 #include "http.h"
-
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <filesystem>
 #include <fstream>
@@ -15,8 +13,6 @@
 #include <string_view>
 #include <stdexcept>
 
-namespace nl = nlohmann;
-
 #if defined(_WIN32)
 #define WIN32_LEAN_AND_MEAN
 #ifndef NOMINMAX
@@ -195,8 +191,8 @@ static void safe_write_file(const fs::path & path, const std::string & data) {
     }
 }
 
-static nl::json api_get(const std::string & url,
-                        const std::string & token) {
+static common_json api_get(const std::string & url,
+                           const std::string & token) {
     auto [cli, parts] = common_http_client(url);
 
     httplib::Headers headers = {
@@ -214,10 +210,10 @@ static nl::json api_get(const std::string & url,
         auto body = res->body;
 
         if (res->status == 200) {
-            return nl::json::parse(res->body);
+            return common_json::parse(res->body);
         }
         try {
-            body = nl::json::parse(res->body)["error"].get<std::string>();
+            body = common_json::parse(res->body)["error"].get<std::string>();
         } catch (...) { }
 
         throw std::runtime_error("GET failed (" + std::to_string(res->status) + "): " + body);
@@ -280,7 +276,7 @@ static std::string get_repo_commit(const std::string & repo_id,
         safe_write_file(refs_path / name, commit);
         return commit;
 
-    } catch (const nl::json::exception & e) {
+    } catch (const common_json_error & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());
@@ -358,7 +354,7 @@ hf_files get_repo_files(const std::string & repo_id,
 
             files.push_back(file);
         }
-    } catch (const nl::json::exception & e) {
+    } catch (const common_json_error & e) {
         LOG_ERR("%s: JSON error: %s\n", __func__, e.what());
     } catch (const std::exception & e) {
         LOG_ERR("%s: error: %s\n", __func__, e.what());
diff --git a/common/jinja/README.md b/common/jinja/README.md
index 8291240767e8..5b97fc92c5a7 100644
--- a/common/jinja/README.md
+++ b/common/jinja/README.md
@@ -7,7 +7,7 @@ The implementation can be found in the `common/jinja` directory.
 ## Key Features
 
 - Input marking: security against special token injection
-- Decoupled from `nlohmann::json`: this dependency is only used for JSON-to-internal type translation and is completely optional
+- Decoupled from the JSON library: `common_json` is only used for JSON-to-internal type translation and is completely optional
 - Minimal primitive types: int, float, bool, string, array, object, none, undefined
 - Detailed logging: allow source tracing on error
 - Clean architecture: workarounds are applied to input data before entering the runtime (see `common/chat.cpp`)
diff --git a/common/jinja/caps.cpp b/common/jinja/caps.cpp
index 6e3a1e9b298a..9971c021e188 100644
--- a/common/jinja/caps.cpp
+++ b/common/jinja/caps.cpp
@@ -4,14 +4,14 @@
 
 // note: the json dependency is only for defining input in a convenient way
 // we can remove it in the future when we figure out a better way to define inputs using jinja::value
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <functional>
 #include <sstream>
 
 #define FILENAME "jinja-caps"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 namespace jinja {
 
diff --git a/common/jinja/value.cpp b/common/jinja/value.cpp
index 870596d617fb..6999ef7d6706 100644
--- a/common/jinja/value.cpp
+++ b/common/jinja/value.cpp
@@ -3,7 +3,7 @@
 #include "value.h"
 
 // for converting from JSON to jinja values
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <sstream>
 #include <string>
@@ -1355,7 +1355,7 @@ const func_builtins & value_undefined_t::get_builtins() const {
 //////////////////////////////////
 
 
-static value from_json(const nlohmann::ordered_json & j, bool mark_input) {
+static value from_json(const common_json & j, bool mark_input) {
     if (j.is_null()) {
         return mk_val<value_none>();
     } else if (j.is_boolean()) {
@@ -1452,7 +1452,7 @@ bool value_compare(const value & a, const value & b, value_compare_op op) {
 }
 
 template<>
-void global_from_json(context & ctx, const nlohmann::ordered_json & json_obj, bool mark_input) {
+void global_from_json(context & ctx, const common_json & json_obj, bool mark_input) {
     // printf("global_from_json: %s\n" , json_obj.dump(2).c_str());
     if (json_obj.is_null() || !json_obj.is_object()) {
         throw std::runtime_error("global_from_json: input JSON value must be an object");
diff --git a/common/jinja/value.h b/common/jinja/value.h
index 5cf85e4f5443..4926fb68016d 100644
--- a/common/jinja/value.h
+++ b/common/jinja/value.h
@@ -86,7 +86,7 @@ struct context; // forward declaration
 // marking input can be useful for tracking data provenance
 // and preventing template injection attacks
 //
-// Note: T_JSON can be nlohmann::ordered_json
+// Note: T_JSON can be common_json
 template<typename T_JSON>
 void global_from_json(context & ctx, const T_JSON & json_obj, bool mark_input);
 
diff --git a/common/json-schema-to-grammar.cpp b/common/json-schema-to-grammar.cpp
index 955b4e014b09..0aee51b26e84 100644
--- a/common/json-schema-to-grammar.cpp
+++ b/common/json-schema-to-grammar.cpp
@@ -1,9 +1,8 @@
 #include "json-schema-to-grammar.h"
 #include "common.h"
 
-#include <nlohmann/json.hpp>
-
 #include <algorithm>
+#include <limits>
 #include <map>
 #include <regex>
 #include <sstream>
@@ -12,7 +11,7 @@
 #include <unordered_set>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static std::string build_repetition(const std::string & item_rule, int min_items, int max_items, const std::string & separator_rule = "") {
     auto has_max = max_items != std::numeric_limits<int>::max();
@@ -917,7 +916,11 @@ class common_schema_converter {
             return _add_rule(rule_name, _resolve_ref(schema["$ref"]));
         }
         if (schema.contains("oneOf") || schema.contains("anyOf")) {
-            std::vector<json> alt_schemas = schema.contains("oneOf") ? schema["oneOf"].get<std::vector<json>>() : schema["anyOf"].get<std::vector<json>>();
+            const json & alts = schema.contains("oneOf") ? schema.at("oneOf") : schema.at("anyOf");
+            std::vector<json> alt_schemas;
+            for (const auto & alt : alts) {
+                alt_schemas.push_back(alt);
+            }
             return _add_rule(rule_name, _generate_union_rule(name, alt_schemas));
         }
         if (schema_type.is_array()) {
@@ -1111,7 +1114,7 @@ common_schema_info::~common_schema_info() = default;
 common_schema_info::common_schema_info(common_schema_info &&) noexcept = default;
 common_schema_info & common_schema_info::operator=(common_schema_info &&) noexcept = default;
 
-void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
+void common_schema_info::resolve_refs(common_json & schema) {
     impl_->resolve_refs(schema, "");
 }
 
@@ -1119,7 +1122,7 @@ void common_schema_info::resolve_refs(nlohmann::ordered_json & schema) {
 // Some models emit raw string values rather than JSON-encoded strings for string parameters.
 // If any branch of the schema (via oneOf, anyOf, $ref, etc.) permits a string, this returns
 // true, allowing callers to handle the value as a raw string for simplicity.
-bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schema) {
+bool common_schema_info::resolves_to_string(const common_json & schema) {
     std::unordered_set<std::string> visited_refs;
 
     std::function<bool(const json &)> check = [&](const json & s) -> bool {
@@ -1227,7 +1230,7 @@ bool common_schema_info::resolves_to_string(const nlohmann::ordered_json & schem
     return check(schema);
 }
 
-std::string json_schema_to_grammar(const json & schema, bool force_gbnf) {
+std::string json_schema_to_grammar(const common_json & schema, bool force_gbnf) {
 #ifdef LLAMA_USE_LLGUIDANCE
     if (!force_gbnf) {
         return "%llguidance {}\nstart: %json " + schema.dump();
@@ -1248,10 +1251,10 @@ std::string build_grammar(const std::function<void(const common_grammar_builder
         /* .add_rule = */ [&](const std::string & name, const std::string & rule) {
             return converter._add_rule(name, rule);
         },
-        /* .add_schema = */ [&](const std::string & name, const nlohmann::ordered_json & schema) {
+        /* .add_schema = */ [&](const std::string & name, const common_json & schema) {
             return converter.visit(schema, name == "root" ? "" : name);
         },
-        /* .resolve_refs = */ [&](nlohmann::ordered_json & schema) {
+        /* .resolve_refs = */ [&](common_json & schema) {
             converter.resolve_refs(schema, "");
         }
     };
diff --git a/common/json-schema-to-grammar.h b/common/json-schema-to-grammar.h
index 240d64231154..84ed71c76a13 100644
--- a/common/json-schema-to-grammar.h
+++ b/common/json-schema-to-grammar.h
@@ -1,12 +1,12 @@
 #pragma once
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
 #include <functional>
 #include <memory>
 #include <string>
 
-std::string json_schema_to_grammar(const nlohmann::ordered_json & schema,
+std::string json_schema_to_grammar(const common_json & schema,
                                    bool force_gbnf = false);
 
 class common_schema_converter;
@@ -24,14 +24,14 @@ class common_schema_info {
     common_schema_info(common_schema_info &&) noexcept;
     common_schema_info & operator=(common_schema_info &&) noexcept;
 
-    void resolve_refs(nlohmann::ordered_json & schema);
-    bool resolves_to_string(const nlohmann::ordered_json & schema);
+    void resolve_refs(common_json & schema);
+    bool resolves_to_string(const common_json & schema);
 };
 
 struct common_grammar_builder {
     std::function<std::string(const std::string &, const std::string &)> add_rule;
-    std::function<std::string(const std::string &, const nlohmann::ordered_json &)> add_schema;
-    std::function<void(nlohmann::ordered_json &)> resolve_refs;
+    std::function<std::string(const std::string &, const common_json &)> add_schema;
+    std::function<void(common_json &)> resolve_refs;
 };
 
 struct common_grammar_options {
diff --git a/common/json.cpp b/common/json.cpp
new file mode 100644
index 000000000000..547542bb7d74
--- /dev/null
+++ b/common/json.cpp
@@ -0,0 +1,437 @@
+#include "json.h"
+
+#include "ggml.h"
+
+#define JSON_ASSERT GGML_ASSERT
+#include <nlohmann/json.hpp>
+
+#include <iterator>
+#include <new>
+#include <set>
+#include <unordered_map>
+#include <vector>
+
+using nlohmann::ordered_json;
+
+// a common_json is the backing value, so any value of a tree can be used as a common_json
+static_assert(sizeof(ordered_json)  <= sizeof(common_json),  "common_json storage is too small");
+static_assert(alignof(ordered_json) <= alignof(common_json), "common_json alignment is too weak");
+
+// runs fn and gives every error of the backing library as a common_json_error
+template <typename F>
+static decltype(auto) guard(F && fn) {
+    try {
+        return fn();
+    } catch (const ordered_json::exception & e) {
+        throw common_json_error(e.what());
+    }
+}
+
+static ordered_json & as_json(common_json * self) {
+    return *reinterpret_cast<ordered_json *>(self);
+}
+
+static const ordered_json & as_json(const common_json * self) {
+    return *reinterpret_cast<const ordered_json *>(self);
+}
+
+static common_json & as_common(ordered_json & json) {
+    return *reinterpret_cast<common_json *>(&json);
+}
+
+static const common_json & as_common(const ordered_json & json) {
+    return *reinterpret_cast<const common_json *>(&json);
+}
+
+static ordered_json to_json(const common_json_value & val) {
+    switch (val.type) {
+        case common_json_value::VAL_NULL:   return nullptr;
+        case common_json_value::VAL_BOOL:   return val.val_bool;
+        case common_json_value::VAL_INT:    return val.val_int;
+        case common_json_value::VAL_UINT:   return val.val_uint;
+        case common_json_value::VAL_DOUBLE: return val.val_double;
+        case common_json_value::VAL_STRING: return val.val_string;
+        case common_json_value::VAL_JSON:
+            // one owner means no one else can see this tree, so it is safe to move it out
+            // note: this makes a value single use, same as the json_ref of the backing library
+            if (val.val_json.use_count() == 1) {
+                return std::move(as_json(val.val_json.get()));
+            }
+            return as_json(val.val_json.get());
+    }
+
+    return nullptr;
+}
+
+common_json_value::common_json_value(const char * val) {
+    if (val) {
+        type       = VAL_STRING;
+        val_string = val;
+    } else {
+        type = VAL_NULL;
+    }
+}
+
+common_json_value::common_json_value(const common_json & val) :
+    type(VAL_JSON), val_json(std::make_shared<common_json>(val)) {}
+
+common_json_value::common_json_value(common_json && val) :
+    type(VAL_JSON), val_json(std::make_shared<common_json>(std::move(val))) {}
+
+template <typename T>
+common_json_value::common_json_value(const std::set<T> & vals) : type(VAL_JSON) {
+    common_json out = common_json::array();
+
+    for (const auto & val : vals) {
+        out.push_back(val);
+    }
+
+    val_json = std::make_shared<common_json>(std::move(out));
+}
+
+// a set value is usable only for the types below
+#define COMMON_JSON_SET(...) template common_json_value::common_json_value(const std::set<__VA_ARGS__> &);
+
+COMMON_JSON_SET(int)
+COMMON_JSON_SET(std::string)
+
+#undef COMMON_JSON_SET
+
+template <typename T>
+common_json_value::common_json_value(const std::map<std::string, T> & vals) : type(VAL_JSON) {
+    common_json out = common_json::object();
+
+    for (const auto & val : vals) {
+        out.set({ val.first, val.second });
+    }
+
+    val_json = std::make_shared<common_json>(std::move(out));
+}
+
+// a map value is usable only for the types below
+#define COMMON_JSON_MAP(...) template common_json_value::common_json_value(const std::map<std::string, __VA_ARGS__> &);
+
+COMMON_JSON_MAP(bool)
+COMMON_JSON_MAP(std::string)
+
+#undef COMMON_JSON_MAP
+
+template <typename T>
+common_json_value::common_json_value(const std::unordered_map<std::string, T> & vals) : type(VAL_JSON) {
+    common_json out = common_json::object();
+
+    for (const auto & val : vals) {
+        out.set({ val.first, val.second });
+    }
+
+    val_json = std::make_shared<common_json>(std::move(out));
+}
+
+// an unordered map value is usable only for the types below
+#define COMMON_JSON_UMAP(...) template common_json_value::common_json_value(const std::unordered_map<std::string, __VA_ARGS__> &);
+
+COMMON_JSON_UMAP(size_t)
+
+#undef COMMON_JSON_UMAP
+
+template <typename T>
+common_json_value::common_json_value(const std::vector<T> & vals) : type(VAL_JSON) {
+    common_json out = common_json::array();
+
+    for (const auto & val : vals) {
+        out.push_back(val);
+    }
+
+    val_json = std::make_shared<common_json>(std::move(out));
+}
+
+// a vector value is usable only for the types below
+// note: std::vector<bool> is not here, its proxy reference does not convert
+#define COMMON_JSON_VEC(...) template common_json_value::common_json_value(const std::vector<__VA_ARGS__> &);
+
+COMMON_JSON_VEC(int)
+COMMON_JSON_VEC(unsigned char)
+COMMON_JSON_VEC(unsigned int)
+COMMON_JSON_VEC(long)
+COMMON_JSON_VEC(unsigned long)
+COMMON_JSON_VEC(long long)
+COMMON_JSON_VEC(unsigned long long)
+COMMON_JSON_VEC(float)
+COMMON_JSON_VEC(double)
+COMMON_JSON_VEC(std::string)
+COMMON_JSON_VEC(std::vector<float>)
+COMMON_JSON_VEC(common_json)
+
+#undef COMMON_JSON_VEC
+
+common_json_value::common_json_value(std::initializer_list<common_json_item> items) :
+    type(VAL_JSON), val_json(std::make_shared<common_json>(items)) {}
+
+// null, same as the backing library
+// operator[] turns it into an object, push_back() into an array
+common_json::common_json() {
+    new (storage) ordered_json();
+}
+
+common_json::common_json(const common_json & other) {
+    new (storage) ordered_json(as_json(&other));
+}
+
+common_json::common_json(common_json && other) noexcept {
+    new (storage) ordered_json(std::move(as_json(&other)));
+}
+
+common_json::common_json(std::initializer_list<common_json_item> items) {
+    new (storage) ordered_json(ordered_json::object());
+
+    for (const auto & item : items) {
+        set(item);
+    }
+}
+
+common_json::common_json(const common_json_value & val) {
+    new (storage) ordered_json(to_json(val));
+}
+
+common_json::common_json(std::nullptr_t) {
+    new (storage) ordered_json(nullptr);
+}
+
+common_json & common_json::operator=(common_json other) noexcept {
+    as_json(this).swap(as_json(&other));
+
+    return *this;
+}
+
+common_json::~common_json() {
+    as_json(this).~basic_json();
+}
+
+common_json common_json::parse(const std::string & text) {
+    try {
+        // the assignment moves the parsed tree in, it does not copy
+        common_json out;
+        as_json(&out) = ordered_json::parse(text);
+        return out;
+    } catch (const std::exception & e) {
+        throw common_json_error(e.what());
+    }
+}
+
+common_json common_json::parse_no_throw(const std::string & text) {
+    common_json out;
+    as_json(&out) = ordered_json::parse(text, nullptr, false);
+    return out;
+}
+
+bool common_json::is_discarded() const {
+    return as_json(this).is_discarded();
+}
+
+common_json common_json::array() {
+    common_json out;
+    as_json(&out) = ordered_json::array();
+    return out;
+}
+
+common_json common_json::array(std::initializer_list<common_json_value> vals) {
+    common_json out;
+    ordered_json & arr = as_json(&out);
+    arr = ordered_json::array();
+
+    for (const auto & val : vals) {
+        arr.push_back(to_json(val));
+    }
+
+    return out;
+}
+
+common_json common_json::object() {
+    common_json out;
+    as_json(&out) = ordered_json::object();
+    return out;
+}
+
+common_json common_json::object(std::initializer_list<common_json_item> items) {
+    return common_json(items);
+}
+
+common_json common_json::make(const common_json_value & val) {
+    return common_json(val);
+}
+
+bool common_json::is_null()           const { return as_json(this).is_null(); }
+bool common_json::is_object()         const { return as_json(this).is_object(); }
+bool common_json::is_array()          const { return as_json(this).is_array(); }
+bool common_json::is_string()         const { return as_json(this).is_string(); }
+bool common_json::is_boolean()        const { return as_json(this).is_boolean(); }
+bool common_json::is_number()         const { return as_json(this).is_number(); }
+bool common_json::is_number_integer() const { return as_json(this).is_number_integer(); }
+bool common_json::is_number_float()   const { return as_json(this).is_number_float(); }
+
+bool   common_json::empty() const { return as_json(this).empty(); }
+size_t common_json::size()  const { return as_json(this).size(); }
+
+bool common_json::contains(const std::string & key) const {
+    return as_json(this).contains(key);
+}
+
+bool common_json::operator==(const common_json_value & val) const {
+    // compare a tree in place, to_json() would copy it
+    if (val.type == common_json_value::VAL_JSON) {
+        return as_json(this) == as_json(val.val_json.get());
+    }
+    return as_json(this) == to_json(val);
+}
+
+bool common_json::operator!=(const common_json_value & val) const {
+    return !(*this == val);
+}
+
+common_json       & common_json::at(const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this).at(key)); }); }
+const common_json & common_json::at(const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
+common_json       & common_json::at(size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this).at(idx)); }); }
+const common_json & common_json::at(size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
+
+common_json       & common_json::operator[](const std::string & key)       { return guard([&]() -> common_json       & { return as_common(as_json(this)[key]); }); }
+const common_json & common_json::operator[](const std::string & key) const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(key)); }); }
+common_json       & common_json::operator[](size_t idx)                    { return guard([&]() -> common_json       & { return as_common(as_json(this)[idx]); }); }
+const common_json & common_json::operator[](size_t idx)              const { return guard([&]() -> const common_json & { return as_common(as_json(this).at(idx)); }); }
+
+common_json       & common_json::front()       { return as_common(as_json(this).front()); }
+const common_json & common_json::front() const { return as_common(as_json(this).front()); }
+common_json       & common_json::back()        { return as_common(as_json(this).back()); }
+const common_json & common_json::back()  const { return as_common(as_json(this).back()); }
+
+void common_json::clear() {
+    as_json(this).clear();
+}
+
+void common_json::erase(const std::string & key) {
+    guard([&] { as_json(this).erase(key); });
+}
+
+void common_json::erase(size_t idx) {
+    guard([&] { as_json(this).erase(idx); });
+}
+
+void common_json::assign(const common_json_value & val) {
+    as_json(this) = to_json(val);
+}
+
+void common_json::set(const common_json_item & item) {
+    guard([&] { as_json(this)[item.key] = to_json(item.val); });
+}
+
+void common_json::push_back(const common_json_value & val) {
+    guard([&] { as_json(this).push_back(to_json(val)); });
+}
+
+void common_json::push_back(std::initializer_list<common_json_item> items) {
+    common_json val(items);
+
+    guard([&] { as_json(this).push_back(std::move(as_json(&val))); });
+}
+
+size_t common_json::count(const std::string & key) const {
+    return as_json(this).count(key);
+}
+
+void common_json::insert(const common_json & vals) {
+    guard([&] {
+        ordered_json & self = as_json(this);
+
+        self.insert(self.end(), as_json(&vals).begin(), as_json(&vals).end());
+    });
+}
+
+std::string common_json::dump(int indent) const {
+    return guard([&] { return as_json(this).dump(indent); });
+}
+
+std::string common_json::dump_safe(int indent) const {
+    return as_json(this).dump(indent, ' ', false, ordered_json::error_handler_t::replace);
+}
+
+// an array is indexed directly, an object needs a walk from the start
+common_json & common_json::iterator::operator*() const {
+    return guard([&]() -> common_json & {
+        ordered_json & j = as_json(node);
+
+        if (j.is_object()) {
+            return as_common(std::next(j.begin(), idx).value());
+        }
+        if (j.is_array()) {
+            return as_common(j[idx]);
+        }
+
+        // a plain value gives itself once, same as the backing library
+        return *node;
+    });
+}
+
+std::string common_json::iterator::key() const {
+    return guard([&] { return std::next(as_json(node).begin(), idx).key(); });
+}
+
+common_json::iterator common_json::begin() const {
+    return iterator(const_cast<common_json *>(this), 0);
+}
+
+common_json::iterator common_json::end() const {
+    return iterator(const_cast<common_json *>(this), size());
+}
+
+// the keys follow the backing library: the index for an array, "" for a plain value
+common_json::items_view::entry common_json::items_view::iterator::operator*() const {
+    return guard([&]() -> entry {
+        ordered_json & j = as_json(node);
+
+        if (j.is_object()) {
+            auto it = std::next(j.begin(), idx);
+
+            return { it.key(), as_common(it.value()) };
+        }
+        if (j.is_array()) {
+            return { std::to_string(idx), as_common(j[idx]) };
+        }
+
+        return { std::string(), *node };
+    });
+}
+
+common_json::items_view common_json::items() const {
+    return items_view(const_cast<common_json *>(this), size());
+}
+
+template <typename T> T common_json::get() const {
+    return guard([&] { return as_json(this).get<T>(); });
+}
+
+// the backing library cannot build a common_json, so this one is just a copy
+template <> common_json common_json::get<common_json>() const {
+    return *this;
+}
+
+// get<T>() is usable only for the types below
+
+#define COMMON_JSON_GET(...) template __VA_ARGS__ common_json::get<__VA_ARGS__>() const;
+
+COMMON_JSON_GET(bool)
+COMMON_JSON_GET(int)
+COMMON_JSON_GET(unsigned int)
+COMMON_JSON_GET(long)
+COMMON_JSON_GET(unsigned long)
+COMMON_JSON_GET(long long)
+COMMON_JSON_GET(unsigned long long)
+COMMON_JSON_GET(float)
+COMMON_JSON_GET(double)
+COMMON_JSON_GET(std::string)
+COMMON_JSON_GET(std::vector<float>)
+COMMON_JSON_GET(std::vector<std::string>)
+COMMON_JSON_GET(std::set<std::string>)
+COMMON_JSON_GET(std::vector<int>)
+COMMON_JSON_GET(std::vector<size_t>)
+COMMON_JSON_GET(std::unordered_map<std::string, size_t>)
+
+#undef COMMON_JSON_GET
diff --git a/common/json.h b/common/json.h
new file mode 100644
index 000000000000..9e20a2adb2f0
--- /dev/null
+++ b/common/json.h
@@ -0,0 +1,354 @@
+#pragma once
+
+#include <cstddef>
+#include <cstdint>
+#include <initializer_list>
+#include <iterator>
+#include <map>
+#include <memory>
+#include <set>
+#include <stdexcept>
+#include <string>
+#include <string_view>
+#include <type_traits>
+#include <unordered_map>
+#include <utility>
+#include <vector>
+
+// common_json, a thin wrapper around vendor json library
+// the underlay library is pimpl, we are using nlohmann::json for now
+//
+// many features of the library are deliberately left out, to keep this interface small and generic and to keep compile time down
+//
+// some main differences compared to nlohmann::json :
+// - object keys keep the order in which they are added
+// - errors are always throw as common_json_error
+// - obj.push_back({key, val}) is intentionally unsupported to avoid confusion with push_back on a vector; write it as obj[key] = val for clarity
+// - a braced pair in value position does not build, e.g. {"key", {"a", "b"}}; write array({"a", "b"}) where nlohmann made an array
+//
+// in doubt, search the code base for an existing usage example; do not add anything to this header unless absolutely necessary
+
+class common_json;
+
+// common_json_value holds a list of these, and each of them holds a value, so one must come first
+struct common_json_item;
+
+struct common_json_error : std::runtime_error {
+    using std::runtime_error::runtime_error;
+};
+
+// one value, tagged so that this header stays free of the backing library
+// note: a value that holds a tree is single use, the second use gives null
+struct common_json_value {
+    enum value_type {
+        VAL_NULL,
+        VAL_BOOL,
+        VAL_INT,
+        VAL_UINT,
+        VAL_DOUBLE,
+        VAL_STRING,
+        VAL_JSON,
+    };
+
+    value_type type = VAL_NULL;
+
+    union {
+        bool     val_bool;
+        int64_t  val_int;
+        uint64_t val_uint = 0;
+        double   val_double;
+    };
+
+    std::string                  val_string;
+    std::shared_ptr<common_json> val_json;
+
+    common_json_value(std::nullptr_t = nullptr) : type(VAL_NULL) {}
+    common_json_value(bool val) : type(VAL_BOOL), val_bool(val) {}
+    common_json_value(std::string val) : type(VAL_STRING), val_string(std::move(val)) {}
+    // without this a string_view lands on the common_json ctor below and recurses
+    common_json_value(std::string_view val) : type(VAL_STRING), val_string(val) {}
+    common_json_value(const char * val);
+    common_json_value(const common_json & val);
+    common_json_value(common_json && val);
+    // only for the types instantiated in json.cpp, the rest fails at link time
+    template <typename T> common_json_value(const std::vector<T> & vals);
+    // a set becomes an array, in the set's own order
+    template <typename T> common_json_value(const std::set<T> & vals);
+    // a map becomes an object, keyed in the map's own order
+    template <typename T> common_json_value(const std::map<std::string, T> & vals);
+    template <typename T> common_json_value(const std::unordered_map<std::string, T> & vals);
+
+    // nested object, e.g. {"fn", {{"name", "x"}}}
+    // note: a nested pair {"a", "b"} does not build, use common_json::array({"a", "b"}) for an array
+    common_json_value(std::initializer_list<common_json_item> items);
+
+    template <typename T, typename std::enable_if<std::is_integral<T>::value && !std::is_same<T, bool>::value, int>::type = 0>
+    common_json_value(T val) : type(std::is_signed<T>::value ? VAL_INT : VAL_UINT) {
+        if (std::is_signed<T>::value) {
+            val_int = (int64_t) val;
+        } else {
+            val_uint = (uint64_t) val;
+        }
+    }
+
+    template <typename T, typename std::enable_if<std::is_floating_point<T>::value, int>::type = 0>
+    common_json_value(T val) : type(VAL_DOUBLE), val_double((double) val) {}
+};
+
+struct common_json_item {
+    std::string       key;
+    common_json_value val;
+
+    template <typename T>
+    common_json_item(std::string key, T && val) :
+        key(std::move(key)), val(std::forward<T>(val)) {}
+
+    // a braced list cannot deduce T, so it needs its own overload
+    common_json_item(std::string key, std::initializer_list<common_json_item> items) :
+        key(std::move(key)), val(items) {}
+};
+
+// the types common_json_value holds on its own
+// anything else reaches its common_json ctor and recurses forever
+template <typename T> struct common_json_is_value : std::integral_constant<bool,
+    std::is_arithmetic<T>::value ||
+    std::is_same<T, std::nullptr_t>::value ||
+    std::is_same<T, std::string>::value ||
+    std::is_same<T, std::string_view>::value ||
+    std::is_same<T, char *>::value ||
+    std::is_same<T, const char *>::value ||
+    std::is_same<T, common_json>::value> {};
+
+template <typename T, typename A>
+struct common_json_is_value<std::vector<T, A>> : std::true_type {};
+
+template <typename T, typename C, typename A>
+struct common_json_is_value<std::set<T, C, A>> : std::true_type {};
+
+template <typename V, typename C, typename A>
+struct common_json_is_value<std::map<std::string, V, C, A>> : std::true_type {};
+
+template <typename V, typename H, typename E, typename A>
+struct common_json_is_value<std::unordered_map<std::string, V, H, E, A>> : std::true_type {};
+
+class common_json {
+  public:
+    common_json();
+    common_json(const common_json & other);
+    common_json(common_json && other) noexcept;
+    common_json(std::initializer_list<common_json_item> items);
+    common_json(const common_json_value & val);
+
+    // direct, a value would need two conversions in a row
+    common_json(std::nullptr_t);
+
+    // one step, so that "abc" or a vector can go straight into a common_json
+    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value &&
+                                                  !std::is_same<typename std::decay<T>::type, common_json_value>::value, int>::type = 0>
+    common_json(T && val) : common_json(common_json_value(std::forward<T>(val))) {
+        static_assert(common_json_is_value<typename std::decay<T>::type>::value,
+                      "no common_json_value ctor holds this type, add one instead of letting it recurse");
+    }
+
+    // by value, same as the backing library
+    // the right side is copied before the left side can invalidate it, e.g. msg["a"] = msg.at("b")
+    common_json & operator=(common_json other) noexcept;
+
+    ~common_json();
+
+    // throws common_json_error if the text is not valid JSON
+    static common_json parse(const std::string & text);
+
+    // gives a discarded value instead of throwing, check it with is_discarded()
+    static common_json parse_no_throw(const std::string & text);
+
+    bool is_discarded() const;
+
+    static common_json array();
+    static common_json array(std::initializer_list<common_json_value> vals);
+    static common_json object();
+    static common_json object(std::initializer_list<common_json_item> items);
+
+    // holds a single value, e.g. make("abc").dump() gives "\"abc\""
+    static common_json make(const common_json_value & val);
+
+    bool is_null()    const;
+    bool is_object()  const;
+    bool is_array()   const;
+    bool is_string()  const;
+    bool is_boolean() const;
+    bool is_number()  const;
+    bool is_number_integer() const;
+    bool is_number_float()   const;
+
+    bool   empty() const;
+    size_t size()  const;
+
+    bool contains(const std::string & key) const;
+
+    bool operator==(const common_json_value & val) const;
+    bool operator!=(const common_json_value & val) const;
+
+    // at() throws common_json_error if the key is missing, operator[] adds a null value instead
+    // note: a const operator[] cannot add, it throws like at()
+    common_json       & at(const std::string & key);
+    const common_json & at(const std::string & key) const;
+    common_json       & at(size_t idx);
+    const common_json & at(size_t idx) const;
+
+    common_json       & operator[](const std::string & key);
+    const common_json & operator[](const std::string & key) const;
+    common_json       & operator[](const char * key)       { return (*this)[std::string(key)]; }
+    const common_json & operator[](const char * key) const { return (*this)[std::string(key)]; }
+    common_json       & operator[](int idx)       { return (*this)[to_idx(idx)]; }
+    const common_json & operator[](int idx) const { return (*this)[to_idx(idx)]; }
+    common_json       & operator[](size_t idx);
+    const common_json & operator[](size_t idx) const;
+
+    common_json       & front();
+    const common_json & front() const;
+    common_json       & back();
+    const common_json & back()  const;
+
+    void clear();
+
+    void erase(const std::string & key);
+    void erase(size_t idx);
+
+    // only for the types instantiated in json.cpp, the rest fails at link time
+    template <typename T> T get() const;
+
+    // implicit get<T>() for plain values, so they can be assigned to their C++ type directly
+    // note: kept to this short list on purpose, a wider one makes j["key"] ambiguous
+    // note: a numeric one would make "str = json;" ambiguous, a number converts to char too
+    operator std::string() const { return get<std::string>(); }
+
+    template <typename T>
+    T value(const std::string & key, T def) const {
+        return contains(key) ? at(key).get<T>() : def;
+    }
+
+    std::string value(const std::string & key, const char * def) const {
+        return contains(key) ? at(key).get<std::string>() : std::string(def);
+    }
+
+    // a JSON default needs no get<T>(), it is already the right type
+    common_json value(const std::string & key, const common_json & def) const {
+        return contains(key) ? at(key) : def;
+    }
+
+    void assign(const common_json_value & val);
+    void set(const common_json_item & item);
+    void push_back(const common_json_value & val);
+
+    // appends one object, e.g. push_back({{"a", 1}})
+    void push_back(std::initializer_list<common_json_item> items);
+
+    // 1 if the key is there, 0 if not
+    size_t count(const std::string & key) const;
+
+    // appends every value of another array; inserting an array into itself throws
+    void insert(const common_json & vals);
+
+    // a common_json goes through the copy assignment above, everything else becomes a value
+    template <typename T, typename std::enable_if<!std::is_same<typename std::decay<T>::type, common_json>::value, int>::type = 0>
+    common_json & operator=(T && val) {
+        assign(common_json_value(std::forward<T>(val)));
+        return *this;
+    }
+
+    std::string dump(int indent = -1) const;
+
+    // same as dump(), but bad UTF-8 gets replaced instead of throwing
+    std::string dump_safe(int indent = -1) const;
+
+    // walks an array by index, or an object in insertion order
+    // a plain value gives itself once, same as the backing library
+    class iterator {
+      public:
+        using iterator_category = std::forward_iterator_tag;
+        using value_type        = common_json;
+        using difference_type   = std::ptrdiff_t;
+        using pointer           = common_json *;
+        using reference         = common_json &;
+
+        iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
+
+        common_json & operator*() const;
+        common_json & value()     const { return **this; }
+        std::string   key()       const;
+
+        iterator & operator++() {
+            idx++;
+            return *this;
+        }
+
+        bool operator!=(const iterator & other) const { return idx != other.idx; }
+        bool operator==(const iterator & other) const { return idx == other.idx; }
+
+      private:
+        common_json * node;
+        size_t        idx;
+    };
+
+    iterator begin() const;
+    iterator end()   const;
+
+    // allows: for (const auto & [key, val] : obj.items())
+    class items_view {
+      public:
+        // the members are public, so an entry also works with structured bindings
+        struct entry {
+            std::string   k;
+            common_json & v;
+
+            const std::string & key()   const { return k; }
+            common_json &       value() const { return v; }
+        };
+
+        items_view(common_json * node, size_t n) : node(node), n(n) {}
+
+        class iterator {
+          public:
+            iterator(common_json * node, size_t idx) : node(node), idx(idx) {}
+
+            entry operator*() const;
+
+            iterator & operator++() {
+                idx++;
+                return *this;
+            }
+
+            bool operator!=(const iterator & other) const { return idx != other.idx; }
+
+          private:
+            common_json * node;
+            size_t        idx;
+        };
+
+        iterator begin() const { return iterator(node, 0); }
+        iterator end()   const { return iterator(node, n); }
+
+      private:
+        common_json * node;
+        size_t        n;
+    };
+
+    items_view items() const;
+
+  private:
+    // a negative index must not turn into a huge size_t
+    static size_t to_idx(int idx) {
+        if (idx < 0) {
+            throw common_json_error("negative array index");
+        }
+        return (size_t) idx;
+    }
+
+    // the backing value is built here, json.cpp checks that it fits
+    // it cannot be a pointer: a value inside a tree would then not be a common_json
+    // at() could then only give back a copy instead of a real reference
+    alignas(8) unsigned char storage[32];
+};
+
+using common_json_entry = common_json::items_view::entry;
diff --git a/common/peg-parser.cpp b/common/peg-parser.cpp
index 4a4be7cf789f..46fc29bf2f8b 100644
--- a/common/peg-parser.cpp
+++ b/common/peg-parser.cpp
@@ -10,7 +10,6 @@
 #include <initializer_list>
 #include <map>
 #include <memory>
-#include <nlohmann/json.hpp>
 #include <regex>
 #include <set>
 #include <stdexcept>
@@ -1120,8 +1119,8 @@ common_peg_parser common_peg_parser_builder::chars(const std::string & classes,
     return wrap(arena_.add_parser(common_peg_chars_parser{classes, ranges, negated, min, max}));
 }
 
-common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw) {
-    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<nlohmann::ordered_json>(schema), raw}));
+common_peg_parser common_peg_parser_builder::schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw) {
+    return wrap(arena_.add_parser(common_peg_schema_parser{p.id(), name, std::make_shared<common_json>(schema), raw}));
 }
 
 common_peg_parser common_peg_parser_builder::rule(const std::string & name, const common_peg_parser & p, bool trigger) {
@@ -1805,8 +1804,8 @@ void common_peg_arena::build_grammar(const common_grammar_builder & builder, boo
     }
 }
 
-static nlohmann::json serialize_parser_variant(const common_peg_parser_variant & variant) {
-    using json = nlohmann::json;
+static common_json serialize_parser_variant(const common_peg_parser_variant & variant) {
+    using json = common_json;
 
     return std::visit([](const auto & p) -> json {
         using T = std::decay_t<decltype(p)>;
@@ -1860,7 +1859,7 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
                 {"type", "schema"},
                 {"child", p.child},
                 {"name", p.name},
-                {"schema", p.schema ? *p.schema : nullptr},
+                {"schema", p.schema ? *p.schema : json(nullptr)},
                 {"raw", p.raw}
             };
         } else if constexpr (std::is_same_v<T, common_peg_rule_parser>) {
@@ -1888,19 +1887,19 @@ static nlohmann::json serialize_parser_variant(const common_peg_parser_variant &
     }, variant);
 }
 
-nlohmann::json common_peg_arena::to_json() const {
-    auto parsers = nlohmann::json::array();
+common_json common_peg_arena::to_json() const {
+    auto parsers = common_json::array();
     for (const auto & parser : parsers_) {
         parsers.push_back(serialize_parser_variant(parser));
     }
-    return nlohmann::json{
+    return common_json{
         {"parsers", parsers},
         {"rules", rules_},
         {"root", root_}
     };
 }
 
-static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json & j) {
+static common_peg_parser_variant deserialize_parser_variant(const common_json & j) {
     if (!j.contains("type") || !j["type"].is_string()) {
         throw std::runtime_error("Parser variant JSON missing or invalid 'type' field");
     }
@@ -1969,9 +1968,9 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         }
         common_peg_chars_parser parser;
         parser.pattern = j["pattern"];
-        parser.negated = j["negated"];
-        parser.min_count = j["min_count"];
-        parser.max_count = j["max_count"];
+        parser.negated = j["negated"].get<bool>();
+        parser.min_count = j["min_count"].get<int>();
+        parser.max_count = j["max_count"].get<int>();
         for (const auto & range_json : j["ranges"]) {
             if (!range_json.contains("start") || !range_json.contains("end")) {
                 throw std::runtime_error("char_range missing 'start' or 'end' field");
@@ -2007,7 +2006,7 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
         parser.child = j["child"].get<common_peg_parser_id>();
         parser.name = j["name"];
         if (!j["schema"].is_null()) {
-            parser.schema = std::make_shared<nlohmann::ordered_json>(j["schema"]);
+            parser.schema = std::make_shared<common_json>(j["schema"]);
         }
         parser.raw = j["raw"].get<bool>();
         return parser;
@@ -2069,7 +2068,7 @@ static common_peg_parser_variant deserialize_parser_variant(const nlohmann::json
     throw std::runtime_error("Unknown parser type: " + type);
 }
 
-common_peg_arena common_peg_arena::from_json(const nlohmann::json & j) {
+common_peg_arena common_peg_arena::from_json(const common_json & j) {
     if (!j.contains("parsers") || !j["parsers"].is_array()) {
         throw std::runtime_error("JSON missing or invalid 'parsers' array");
     }
@@ -2109,7 +2108,7 @@ std::string common_peg_arena::save() const {
 }
 
 void common_peg_arena::load(const std::string & data) {
-    *this = from_json(nlohmann::json::parse(data));
+    *this = from_json(common_json::parse(data));
 }
 
 common_peg_arena build_peg_parser(const std::function<common_peg_parser(common_peg_parser_builder & builder)> & fn) {
diff --git a/common/peg-parser.h b/common/peg-parser.h
index c198499dd934..ab095cc7d671 100644
--- a/common/peg-parser.h
+++ b/common/peg-parser.h
@@ -1,6 +1,6 @@
 #pragma once
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
 #include <memory>
 #include <set>
@@ -245,7 +245,7 @@ struct common_peg_until_parser {
 struct common_peg_schema_parser {
     common_peg_parser_id child;
     std::string name;
-    std::shared_ptr<nlohmann::ordered_json> schema;
+    std::shared_ptr<common_json> schema;
 
     // Indicates if the GBNF should accept a raw string that matches the schema.
     bool raw;
@@ -332,8 +332,8 @@ class common_peg_arena {
 
     std::string dump(common_peg_parser_id id) const;
 
-    nlohmann::json to_json() const;
-    static common_peg_arena from_json(const nlohmann::json & j);
+    common_json to_json() const;
+    static common_peg_arena from_json(const common_json & j);
 
     std::string save() const;
     void load(const std::string & data);
@@ -490,7 +490,7 @@ class common_peg_parser_builder {
 
     // Wraps a parser with JSON schema metadata for grammar generation.
     // Used internally to convert JSON schemas to GBNF grammar rules.
-    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const nlohmann::ordered_json & schema, bool raw = false);
+    common_peg_parser schema(const common_peg_parser & p, const std::string & name, const common_json & schema, bool raw = false);
 
     // Creates a named rule, stores it in the grammar, and returns a ref.
     // If trigger=true, marks this rule as an entry point for lazy grammar generation.
diff --git a/tests/peg-parser/test-json-serialization.cpp b/tests/peg-parser/test-json-serialization.cpp
index a85801060c0b..da63a23bf218 100644
--- a/tests/peg-parser/test-json-serialization.cpp
+++ b/tests/peg-parser/test-json-serialization.cpp
@@ -8,7 +8,7 @@ void test_json_serialization(testing &t) {
     auto json_serialized = original.to_json().dump();
 
     t.test("compare before/after", [&](testing &t) {
-        auto deserialized = common_peg_arena::from_json(nlohmann::json::parse(json_serialized));
+        auto deserialized = common_peg_arena::from_json(common_json::parse(json_serialized));
 
         // Test complex JSON
         std::string input = R"({"name": "test", "values": [1, 2, 3], "nested": {"a": true}})";
@@ -23,6 +23,6 @@ void test_json_serialization(testing &t) {
     });
 
     t.bench("deserialize", [&]() {
-        auto deserialized = common_peg_arena::from_json(nlohmann::json::parse(json_serialized));
+        auto deserialized = common_peg_arena::from_json(common_json::parse(json_serialized));
     }, 100);
 }
diff --git a/tests/peg-parser/tests.h b/tests/peg-parser/tests.h
index debd4286c50a..00e81815b621 100644
--- a/tests/peg-parser/tests.h
+++ b/tests/peg-parser/tests.h
@@ -1,7 +1,7 @@
 #pragma once
 
 // Common includes for all test files
-#include <nlohmann/json.hpp>
+#include "json.h"
 #include <string>
 #include <vector>
 
@@ -11,9 +11,9 @@
 #include "simple-tokenize.h"
 
 struct bench_tool_call {
-    std::string            id;
-    std::string            name;
-    nlohmann::ordered_json args;
+    std::string id;
+    std::string name;
+    common_json args;
 };
 
 // Test function declarations
diff --git a/tests/test-chat-peg-parser.cpp b/tests/test-chat-peg-parser.cpp
index 3ab7a67b6a82..793891394ce6 100644
--- a/tests/test-chat-peg-parser.cpp
+++ b/tests/test-chat-peg-parser.cpp
@@ -11,9 +11,9 @@
 #include <regex>
 #include <string>
 
-#include "nlohmann/json.hpp"
+#include "json.h"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static json create_tools();
 static void test_example_native(testing & t);
@@ -63,10 +63,10 @@ static json create_tools() {
                           { { "type", "string" }, { "description", "The city and state, e.g. San Francisco, CA" } } },
                         { "unit",
                           { { "type", "string" },
-                            { "enum", { "celsius", "fahrenheit" } },
+                            { "enum", json::array({ "celsius", "fahrenheit" }) },
                             { "description",
                               "The temperature unit to use. Infer this from the users location." } } } } },
-                    { "required", { "location", "unit" } },
+                    { "required", json::array({ "location", "unit" }) },
                 } },
           }                      }
     };
@@ -86,14 +86,14 @@ static json create_tools() {
                           { { "type", "string" }, { "description", "The city and state, e.g. San Francisco, CA" } } },
                         { "unit",
                           { { "type", "string" },
-                            { "enum", { "celsius", "fahrenheit" } },
+                            { "enum", json::array({ "celsius", "fahrenheit" }) },
                             { "description", "The temperature unit to use. Infer this from the users location." } } },
                         { "days",
                           { { "type", "integer" },
                             { "description", "Number of days to forecast (1-10)" },
                             { "minimum", 1 },
                             { "maximum", 10 } } } } },
-                    { "required", { "location", "unit" } },
+                    { "required", json::array({ "location", "unit" }) },
                 } },
           }                      }
     };
@@ -114,9 +114,9 @@ static json create_tools() {
                         { "default", 5 } } },
                     { "category",
                       { { "type", "string" },
-                        { "enum", { "api", "troubleshooting", "billing", "general" } },
+                        { "enum", json::array({ "api", "troubleshooting", "billing", "general" }) },
                         { "description", "Filter search by specific category." } } } } },
-                { "required", { "query", "category" } },
+                { "required", json::array({ "query", "category" }) },
                 { "additionalProperties", false } } },
             { "strict", true } } }
     };
@@ -341,7 +341,7 @@ static void test_example_native(testing & t) {
                 { { "invoice_number", { { "type", "string" } } },
                   { "amount", { { "type", "number" } } },
                   { "due_date", { { "type", "string" } } } } },
-              { "required", { "invoice_number", "amount", "due_date" } } },
+              { "required", json::array({ "invoice_number", "amount", "due_date" }) } },
          /* .parallel_tool_calls =  */ false,
          /* .generation_prompt =    */ "<think>",
          /* .input =                */
@@ -406,7 +406,7 @@ static void test_example_qwen3_coder(testing & t) {
 
             std::set<std::string> required_properties;
             if (function.contains("required")) {
-                function.at("required").get_to(required_properties);
+                required_properties = function.at("required").get<std::set<std::string>>();
             }
 
             std::vector<common_peg_parser> arg_parsers;
@@ -661,8 +661,8 @@ void test_command7_parser_compare(testing & t) {
         "5. Provide a detailed cost breakdown that includes accommodation, transportation, meals, and entry fees "
         "to attractions.";
 
-    std::vector<std::tuple<std::string, std::string, nlohmann::json>> tool_calls = {
-        { "call_0", "plan_trip", nlohmann::json::parse(R"({
+    std::vector<std::tuple<std::string, std::string, common_json>> tool_calls = {
+        { "call_0", "plan_trip", common_json::parse(R"({
             "destination": "Japan",
             "duration": 14,
             "budget": 4000,
@@ -686,16 +686,16 @@ void test_command7_parser_compare(testing & t) {
     if (!tool_calls.empty()) {
         tokens.emplace_back("<|START_ACTION|>");
 
-        auto json = nlohmann::json::array();
+        auto json = common_json::array();
         for (const auto & tc : tool_calls) {
-            auto tc_json            = nlohmann::json::object();
+            auto tc_json            = common_json::object();
             tc_json["tool_call_id"] = std::get<0>(tc);
             tc_json["tool_name"]    = std::get<1>(tc);
             tc_json["parameters"]   = std::get<2>(tc);
             json.push_back(tc_json);
         }
 
-        auto tokenized = simple_tokenize(json.dump(-1, ' ', true));
+        auto tokenized = simple_tokenize(json.dump(-1));
         tokens.insert(tokens.end(), tokenized.begin(), tokenized.end());
 
         tokens.emplace_back("<|END_ACTION|>");
@@ -737,7 +737,7 @@ static void test_prefix_tool_names(testing & t) {
                       {
                           { "arg1", { { "type", "integer" } } },
                       } },
-                    { "required", { "arg1" } },
+                    { "required", json::array({ "arg1" }) },
                 } },
           } }
     };
@@ -757,7 +757,7 @@ static void test_prefix_tool_names(testing & t) {
                           { "arg1", { { "type", "integer" } } },
                           { "arg2", { { "type", "integer" } } },
                       } },
-                    { "required", { "arg1" } },
+                    { "required", json::array({ "arg1" }) },
                 } },
           } }
     };
diff --git a/tests/test-chat-template.cpp b/tests/test-chat-template.cpp
index 6a6292cd0151..bcc574afe981 100644
--- a/tests/test-chat-template.cpp
+++ b/tests/test-chat-template.cpp
@@ -7,7 +7,7 @@
 #include <fstream>
 #include <filesystem>
 
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #undef NDEBUG
 #include <cassert>
@@ -20,7 +20,7 @@
 #include "jinja/lexer.h"
 #include "jinja/caps.h"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static int main_automated_tests(void);
 
@@ -304,8 +304,8 @@ void run_single(const std::string& contents, json input, bool use_common, bool d
         if (input.contains("eos_token")) {
             eos_token = input["eos_token"].get<std::string>();
         }
-        nlohmann::ordered_json msgs_json = input["messages"];
-        nlohmann::ordered_json tools_json = input["tools"];
+        common_json msgs_json = input["messages"];
+        common_json tools_json = input["tools"];
         auto messages = common_chat_msgs_parse_oaicompat(msgs_json);
         auto tools = common_chat_tools_parse_oaicompat(tools_json);
         auto output = format_using_common(contents, bos_token, eos_token, messages, tools);
diff --git a/tests/test-chat.cpp b/tests/test-chat.cpp
index c4670da85302..7918f0ffcf48 100644
--- a/tests/test-chat.cpp
+++ b/tests/test-chat.cpp
@@ -19,12 +19,12 @@
 #include <fstream>
 #include <functional>
 #include <iostream>
-#include <nlohmann/json.hpp>
+#include "json.h"
 #include <set>
 #include <stdexcept>
 #include <string>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static std::ostream & operator<<(std::ostream & os, const common_chat_msg_diff & diff) {
     os << "{ content_delta: " << diff.content_delta << "; ";
diff --git a/tests/test-grammar-integration.cpp b/tests/test-grammar-integration.cpp
index 4d5d13dd0d38..eb4b7c78f50f 100644
--- a/tests/test-grammar-integration.cpp
+++ b/tests/test-grammar-integration.cpp
@@ -7,13 +7,13 @@
 #include "../src/unicode.h"
 #include "../src/llama-grammar.h"
 
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <cassert>
 #include <string>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static llama_grammar * build_grammar_with_root(const std::string & grammar_str, const char * grammar_root) {
     return llama_grammar_init_impl(nullptr, grammar_str.c_str(), grammar_root, false, nullptr, 0, nullptr, 0);
diff --git a/tests/test-jinja.cpp b/tests/test-jinja.cpp
index 1eb2a062b75f..974a3f9dd8df 100644
--- a/tests/test-jinja.cpp
+++ b/tests/test-jinja.cpp
@@ -3,7 +3,7 @@
 #include <random>
 #include <cstdlib>
 
-#include <nlohmann/json.hpp>
+#include "json.h"
 #include "subproc.h"
 
 #include "jinja/runtime.h"
@@ -14,7 +14,7 @@
 
 #include "testing.h"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 static void test_template(testing & t, const std::string & name, const std::string & tmpl, const json & vars, const std::string & expect);
 
@@ -240,7 +240,7 @@ static void test_conditionals(testing & t) {
 
     test_template(t, "is undefined key falsy",
         "{{ 'yes' if not y['x'] else 'no' }}",
-        {{"y", {{}}}},
+        {{"y", json::array({nullptr})}},
         "yes"
     );
 
@@ -282,7 +282,7 @@ static void test_conditionals(testing & t) {
 
     test_template(t, "is non-empty object truthy",
         "{{ 'yes' if y else 'no' }}",
-        {{"y", {"x", false}}},
+        {{"y", json::array({"x", false})}},
         "yes"
     );
 
diff --git a/tests/test-json-schema-to-grammar.cpp b/tests/test-json-schema-to-grammar.cpp
index 74b57cf1b669..214dbe1993b8 100755
--- a/tests/test-json-schema-to-grammar.cpp
+++ b/tests/test-json-schema-to-grammar.cpp
@@ -6,7 +6,7 @@
 
 #include "../src/llama-grammar.h"
 
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <cassert>
 #include <fstream>
@@ -1442,7 +1442,7 @@ static void test_resolves_to_string() {
     auto test = [](const std::string & name, const std::string & schema_str, bool expected) {
         fprintf(stderr, "- %s\n", name.c_str());
         common_schema_info info;
-        auto schema = nlohmann::ordered_json::parse(schema_str);
+        auto schema = common_json::parse(schema_str);
         info.resolve_refs(schema);
         bool result = info.resolves_to_string(schema);
         if (result != expected) {
@@ -1517,7 +1517,7 @@ int main() {
 
     test_all("C++", [](const TestCase & tc) {
         try {
-            tc.verify(json_schema_to_grammar(nlohmann::ordered_json::parse(tc.schema), true));
+            tc.verify(json_schema_to_grammar(common_json::parse(tc.schema), true));
             tc.verify_status(SUCCESS);
         } catch (const std::invalid_argument & ex) {
             fprintf(stderr, "Error: %s\n", ex.what());
@@ -1531,7 +1531,7 @@ int main() {
         auto run = [](const TestCase & tc) {
             fprintf(stderr, "- %s\n", tc.name.c_str());
             try {
-                tc.verify(json_schema_to_grammar(nlohmann::ordered_json::parse(tc.schema), true));
+                tc.verify(json_schema_to_grammar(common_json::parse(tc.schema), true));
                 tc.verify_status(SUCCESS);
             } catch (const std::invalid_argument & ex) {
                 fprintf(stderr, "Error: %s\n", ex.what());
diff --git a/tests/test-model-resolution.cpp b/tests/test-model-resolution.cpp
index 2437eeec608c..5191e77514ab 100644
--- a/tests/test-model-resolution.cpp
+++ b/tests/test-model-resolution.cpp
@@ -9,7 +9,7 @@
 #include "http.h"
 #include "log.h"
 
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <algorithm>
 #include <cstdio>
@@ -55,7 +55,7 @@ static const char * COMMIT = "aaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaaa";
 static void serve_repos(httplib::Server & server) {
     server.Get(R"(/api/models/(.+)/refs)", [](const httplib::Request & req, httplib::Response & res) {
         if (g_repos.count(req.matches[1])) {
-            res.set_content(nlohmann::json{{"branches", {{{"name", "main"}, {"targetCommit", COMMIT}}}}}.dump(),
+            res.set_content(common_json{{"branches", common_json::array({ common_json{{"name", "main"}, {"targetCommit", COMMIT}} })}}.dump(),
                             "application/json");
         } else {
             res.status = 404;
@@ -66,7 +66,7 @@ static void serve_repos(httplib::Server & server) {
             res.status = 404;
             return;
         }
-        auto files = nlohmann::json::array();
+        auto files = common_json::array();
         size_t i = 0;
         for (const auto & p : g_repos[req.matches[1]]) {
             char oid[41];
diff --git a/tools/cli/cli-context.cpp b/tools/cli/cli-context.cpp
index 3d801b73d4c1..aa4eb76796d9 100644
--- a/tools/cli/cli-context.cpp
+++ b/tools/cli/cli-context.cpp
@@ -6,8 +6,7 @@
 #include "log.h"
 #include "console.h"
 
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <algorithm>
 #include <cctype>
@@ -16,7 +15,7 @@
 #include <map>
 #include <set>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 struct cli_context_impl {
     json messages      = json::array();
@@ -73,7 +72,7 @@ static std::string format_error_message(const json & err) {
 
 // err is the raw response body of a failed request; it may or may not be JSON
 static std::string format_error_message(const std::string & err) {
-    json parsed = json::parse(err, nullptr, false);
+    json parsed = json::parse_no_throw(err);
     if (!parsed.is_discarded()) {
         return format_error_message(parsed);
     }
@@ -157,7 +156,7 @@ bool cli_context::init() {
             if (!list_and_ask_models()) {
                 return false;
             }
-        } catch (const json::parse_error & e) {
+        } catch (const common_json_error & e) {
             ui::show_error(e.what());
             ui::show_message("This might be caused by an incorrect server-base endpoint URL");
             return false;
@@ -364,7 +363,7 @@ bool cli_context::generate_completion(generated_content & content_out, cli_timin
     ui::assistant_turn a;
 
     std::string err = client.post_sse("/v1/chat/completions", body.dump(), should_stop, [&](const std::string & payload) {
-        json chunk = json::parse(payload, nullptr, false);
+        json chunk = json::parse_no_throw(payload);
         if (chunk.is_discarded()) {
             return;
         }
diff --git a/tools/parser/debug-template-parser.cpp b/tools/parser/debug-template-parser.cpp
index 8a916f79c78e..abe427022332 100644
--- a/tools/parser/debug-template-parser.cpp
+++ b/tools/parser/debug-template-parser.cpp
@@ -5,7 +5,7 @@
 #include "gguf.h"
 #include "jinja/runtime.h"
 #include "log.h"
-#include "nlohmann/json.hpp"
+#include "json.h"
 #include "peg-parser.h"
 
 #include <fstream>
@@ -15,7 +15,7 @@
 #include <sstream>
 #include <string>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 enum class output_mode {
     ANALYSIS,  // Only output analysis results (default)
diff --git a/tools/parser/template-analysis.cpp b/tools/parser/template-analysis.cpp
index bf898a2290f1..11225bd8c000 100644
--- a/tools/parser/template-analysis.cpp
+++ b/tools/parser/template-analysis.cpp
@@ -11,9 +11,9 @@
 #include <vector>
 #include <algorithm>
 
-#include "nlohmann/json.hpp"
+#include "json.h"
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 // ANSI color codes - using 256-color palette for brighter colors (all bold)
 #define ANSI_RESET       "\033[0m"
diff --git a/tools/server/server-chat.cpp b/tools/server/server-chat.cpp
index 0322e54ccea8..a6fe3c6ba619 100644
--- a/tools/server/server-chat.cpp
+++ b/tools/server/server-chat.cpp
@@ -153,7 +153,7 @@ json server_chat_convert_responses_to_chatcmpl(const json & response_body) {
                         prev_msg["content"] = json::array();
                     }
                     auto & prev_content = prev_msg["content"];
-                    prev_content.insert(prev_content.end(), chatcmpl_content.begin(), chatcmpl_content.end());
+                    prev_content.insert(chatcmpl_content);
                 } else {
                     item.erase("status");
                     item.erase("type");
diff --git a/tools/server/server-chat.h b/tools/server/server-chat.h
index 102eae688a31..86b842650ea3 100644
--- a/tools/server/server-chat.h
+++ b/tools/server/server-chat.h
@@ -6,9 +6,7 @@
 #include "server-common.h"
 #include "server-http.h"
 
-#include <nlohmann/json_fwd.hpp>
-
-using json = nlohmann::ordered_json;
+#include "json.h"
 
 // Convert OpenAI Responses API format to OpenAI Chat Completions API format
 json server_chat_convert_responses_to_chatcmpl(const json & body);
diff --git a/tools/server/server-common.cpp b/tools/server/server-common.cpp
index 585f65e83c65..4f5b8202aca5 100644
--- a/tools/server/server-common.cpp
+++ b/tools/server/server-common.cpp
@@ -1540,7 +1540,7 @@ std::vector<llama_token_data> get_token_probabilities(llama_context * ctx, int i
 }
 
 std::string safe_json_to_str(const json & data) {
-    return data.dump(-1, ' ', false, json::error_handler_t::replace);
+    return data.dump_safe();
 }
 
 // TODO: reuse llama_detokenize
diff --git a/tools/server/server-common.h b/tools/server/server-common.h
index 6488be344c6a..f8ea82ef4cf5 100644
--- a/tools/server/server-common.h
+++ b/tools/server/server-common.h
@@ -6,8 +6,7 @@
 #include "chat.h"
 #include "mtmd.h"
 
-#define JSON_ASSERT GGML_ASSERT
-#include <nlohmann/json.hpp>
+#include "json.h"
 
 #include <atomic>
 #include <chrono>
@@ -19,7 +18,7 @@
 #include <string>
 #include <vector>
 
-using json = nlohmann::ordered_json;
+using json = common_json;
 
 #define SLT_DBG(slot, fmt, ...) LOG_DBG("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, ((slot).task ? (slot).task->id : -1), __VA_ARGS__)
 #define SLT_TRC(slot, fmt, ...) LOG_TRC("slot %12.*s: id %2d | task %d | " fmt, 12, __func__, (slot).id, ((slot).task ? (slot).task->id : -1), __VA_ARGS__)
@@ -42,9 +41,9 @@ static T json_value(const json & body, const std::string & key, const T & defaul
     // Fallback null to default value
     if (body.contains(key) && !body.at(key).is_null()) {
         try {
-            return body.at(key);
-        } catch (NLOHMANN_JSON_NAMESPACE::detail::type_error const & err) {
-            LOG_WRN("Wrong type supplied for parameter '%s'. Expected '%s', using default value: %s\n", key.c_str(), json(default_value).type_name(), err.what());
+            return body.at(key).get<T>();
+        } catch (const common_json_error & err) {
+            LOG_WRN("Wrong type supplied for parameter '%s', using default value: %s\n", key.c_str(), err.what());
             return default_value;
         }
     } else {
diff --git a/tools/server/server-context.cpp b/tools/server/server-context.cpp
index 1293c8640267..d7a6b887d868 100644
--- a/tools/server/server-context.cpp
+++ b/tools/server/server-context.cpp
@@ -35,8 +35,6 @@
 #include <windows.h>
 #endif
 
-using json = nlohmann::ordered_json;
-
 constexpr int HTTP_POLLING_SECONDS = 1;
 
 static common_speculative_output_limits server_output_limits(const common_params & params) {
@@ -657,14 +655,14 @@ struct server_slot {
             res["n_prompt_tokens_processed"] = stats.n_prompt_processed;
             res["n_prompt_tokens_cache"]     = stats.n_prompt_cached;
             res["params"] = ptask->params.to_json(only_metrics);
-            res["next_token"] = {
+            res["next_token"] = json::array({
                 {
                     {"has_next_token", has_next_token},
                     {"has_new_line",   has_new_line},
                     {"n_remain",       n_remaining()},
                     {"n_decoded",      stats.n_gen},
                 }
-            };
+            });
 
             if (!only_metrics) {
                 res["prompt"] = ptask->tokens.detokenize(ctx_tgt, true);
@@ -4220,7 +4218,8 @@ std::unique_ptr<server_res_generator> server_routes::handle_completions_impl(
         // tasks.reserve(inputs.size()); // TODO: this is inaccurate due to child tasks
 
         // message delimiters for checkpointing
-        auto delimiters = common_chat_msg_delimiters_parse(json_value(data, "message_delimiters", json::array()));
+        json delims = json_value(data, "message_delimiters", json::array());
+        auto delimiters = common_chat_msg_delimiters_parse(delims);
         delimiters.tokenize(ctx_server.vocab);
 
         for (size_t i = 0; i < inputs.size(); i++) {
@@ -4483,8 +4482,8 @@ static json get_res_model_info(const server_context_meta & meta) {
 static json get_res_models(const server_context_meta & meta) {
     // note: do NOT use ctx_server here, otherwise it's not possible to use this during sleep
 
-    return {
-        {"models", {
+    return json{
+        {"models", json::array({
             {
                 {"name",  meta.model_name},
                 {"model", meta.model_name},
@@ -4493,23 +4492,23 @@ static json get_res_models(const server_context_meta & meta) {
                 {"digest", ""}, // dummy value, llama.cpp does not support managing model file's hash
                 {"type", "model"},
                 {"description", ""},
-                {"tags", {""}},
-                {"capabilities", meta.has_mtmd ? json({"completion","multimodal"}) : json({"completion"})},
+                {"tags", json::array({""})},
+                {"capabilities", meta.has_mtmd ? json::array({"completion","multimodal"}) : json::array({"completion"})},
                 {"parameters", ""},
                 {"details", {
                     {"parent_model", ""},
                     {"format", "gguf"},
                     {"family", ""},
-                    {"families", {""}},
+                    {"families", json::array({""})},
                     {"parameter_size", ""},
                     {"quantization_level", ""}
                 }}
             }
-        }},
+        })},
         {"object", "list"},
-        {"data", {
+        {"data", json::array({
             get_res_model_info(meta),
-        }}
+        })}
     };
 }
 
@@ -5045,7 +5044,7 @@ void server_routes::init_routes() {
 
         std::string content;
         if (body.count("tokens") != 0) {
-            const llama_tokens tokens = body.at("tokens");
+            const llama_tokens tokens = body.at("tokens").get<llama_tokens>();
             content = tokens_to_str(ctx_server.vocab, tokens);
         }
 
@@ -5352,7 +5351,7 @@ std::unique_ptr<server_res_generator> server_routes::handle_embeddings_impl(cons
 
     int embd_normalize = params.embd_normalize;
     if (body.count("embd_normalize") != 0) {
-        embd_normalize = body.at("embd_normalize");
+        embd_normalize = body.at("embd_normalize").get<int>();
         if (meta->pooling_type == LLAMA_POOLING_TYPE_NONE) {
             SRV_DBG("embd_normalize is not supported by pooling type %d, ignoring it\n", meta->pooling_type);
         }
diff --git a/tools/server/server-context.h b/tools/server/server-context.h
index 764df0e08524..5d464b8e8cb7 100644
--- a/tools/server/server-context.h
+++ b/tools/server/server-context.h
@@ -4,7 +4,7 @@
 #include "server-task.h"
 #include "server-queue.h"
 
-#include <nlohmann/json_fwd.hpp>
+#include "json.h"
 
 #include <cstddef>
 #include <memory>
diff --git a/tools/server/server-models.cpp b/tools/server/server-models.cpp
index d60545194188..db0fac99527b 100644
--- a/tools/server/server-models.cpp
+++ b/tools/server/server-models.cpp
@@ -2462,7 +2462,7 @@ server_http_proxy::server_http_proxy(
     bool has_files = !files.empty();
 
     if (has_files) {
-        json form_fields = json::parse(body, nullptr, false);
+        json form_fields = json::parse_no_throw(body);
         if (!form_fields.is_discarded()) {
             auto boundary = generate_multipart_boundary();
             effective_body = build_multipart_body(form_fields, files, boundary);
diff --git a/tools/server/server-schema.cpp b/tools/server/server-schema.cpp
index 5d7fa6ae6ec3..64b9251295ce 100644
--- a/tools/server/server-schema.cpp
+++ b/tools/server/server-schema.cpp
@@ -503,7 +503,7 @@ std::vector<std::unique_ptr<field>> make_llama_cmpl_schema(const common_params &
         ->set_handler([&](field_eval_context & ctx, const json & data) {
             const auto & samplers = data.at("samplers");
             if (samplers.is_array()) {
-                ctx.params.sampling.samplers = common_sampler_types_from_names(samplers);
+                ctx.params.sampling.samplers = common_sampler_types_from_names(samplers.get<std::vector<std::string>>());
             } else if (samplers.is_string()) {
                 ctx.params.sampling.samplers = common_sampler_types_from_chars(samplers.get<std::string>());
             }
@@ -580,8 +580,7 @@ static void handle_with_catch(const char * name, std::function<void()> func) {
 
 // treat a null value as absent so clients can send null to request the server default
 static bool has_value(const json & data, const char * n) {
-    auto it = data.find(n);
-    return it != data.end() && !it->is_null();
+    return data.contains(n) && !data.at(n).is_null();
 }
 
 template <typename T>
diff --git a/tools/server/server-task.cpp b/tools/server/server-task.cpp
index 258cdcf8fb86..0d3beb313cea 100644
--- a/tools/server/server-task.cpp
+++ b/tools/server/server-task.cpp
@@ -12,8 +12,6 @@
 
 #include <sstream>
 
-using json = nlohmann::ordered_json;
-
 //
 // task_params
 //
@@ -304,7 +302,7 @@ json completion_token_output::probs_vector_to_json(const std::vector<completion_
 }
 
 float completion_token_output::logarithm(float x) {
-    // nlohmann::json converts -inf to null, so we need to prevent that
+    // the JSON library converts -inf to null, so we need to prevent that
     return x == 0.0f ? std::numeric_limits<float>::lowest() : std::log(x);
 }
 
@@ -407,7 +405,7 @@ json server_task_result_cmpl_final::to_json_oaicompat() {
         res["__verbose"] = to_json_non_oaicompat();
     }
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
 
     return res;
@@ -455,7 +453,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat() {
         res["__verbose"] = to_json_non_oaicompat();
     }
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
 
     return res;
@@ -516,7 +514,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_chat_stream() {
     }
 
     if (stats.is_set()) {
-        deltas.back().push_back({"timings", stats.to_json()});
+        deltas.back()["timings"] = stats.to_json();
     }
 
     // extra fields for debugging purposes
@@ -709,7 +707,7 @@ json server_task_result_cmpl_final::to_json_oaicompat_resp_stream() {
     });
 
     if (stats.is_set()) {
-        server_sent_events.back().at("data").push_back({"timings", stats.to_json()});
+        server_sent_events.back().at("data")["timings"] = stats.to_json();
     }
 
     return server_sent_events;
@@ -1061,10 +1059,10 @@ json server_task_result_cmpl_partial::to_json_non_oaicompat() {
     };
     // populate the timings object when needed (usually for the last response or with timings_per_token enabled)
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
     if (is_progress) {
-        res.push_back({"prompt_progress", progress.to_json()});
+        res["prompt_progress"] = progress.to_json();
     }
     if (!prob_output.probs.empty()) {
         res["completion_probabilities"] = completion_token_output::probs_vector_to_json({prob_output}, post_sampling_probs);
@@ -1101,10 +1099,10 @@ json server_task_result_cmpl_partial::to_json_oaicompat() {
         res["__verbose"] = to_json_non_oaicompat();
     }
     if (stats.is_set()) {
-        res.push_back({"timings", stats.to_json()});
+        res["timings"] = stats.to_json();
     }
     if (is_progress) {
-        res.push_back({"prompt_progress", progress.to_json()});
+        res["prompt_progress"] = progress.to_json();
     }
 
     return res;
@@ -1155,10 +1153,10 @@ json server_task_result_cmpl_partial::to_json_oaicompat_chat() {
         }
 
         if (stats.is_set()) {
-            last_json.push_back({"timings", stats.to_json()});
+            last_json["timings"] = stats.to_json();
         }
         if (is_progress) {
-            last_json.push_back({"prompt_progress", progress.to_json()});
+            last_json["prompt_progress"] = progress.to_json();
         }
     }
 
@@ -1305,10 +1303,10 @@ json server_task_result_cmpl_partial::to_json_oaicompat_resp() {
     if (!events.empty()) {
         json & data = events.back().at("data");
         if (stats.is_set()) {
-            data.push_back({"timings", stats.to_json()});
+            data["timings"] = stats.to_json();
         }
         if (is_progress) {
-            data.push_back({"prompt_progress", progress.to_json()});
+            data["prompt_progress"] = progress.to_json();
         }
     }
 
diff --git a/tools/server/server-task.h b/tools/server/server-task.h
index 25ff01512248..9c99143f8e19 100644
--- a/tools/server/server-task.h
+++ b/tools/server/server-task.h
@@ -11,7 +11,6 @@
 // TODO: prevent including the whole server-common.h as we only use server_tokens
 #include "server-common.h"
 
-using json = nlohmann::ordered_json;
 
 enum server_task_type {
     SERVER_TASK_TYPE_COMPLETION,
diff --git a/tools/server/server-tools.cpp b/tools/server/server-tools.cpp
index b5c5c078ae33..12e9dbb8cfa9 100644
--- a/tools/server/server-tools.cpp
+++ b/tools/server/server-tools.cpp
@@ -2156,7 +2156,7 @@ void server_tools::setup(const std::vector<std::string> & enabled_tools,
                 res->status = 200;
                 res->data   = safe_json_to_str(result);
             }
-        } catch (const json::exception & e) {
+        } catch (const common_json_error & e) {
             res->status = 400;
             res->data   = safe_json_to_str(format_error_response(e.what(), ERROR_TYPE_INVALID_REQUEST));
         } catch (const std::invalid_argument & e) {