Note

This project only worked for the previous major version of Amp. Since Amp is now a server-side product with thin frontends only, this project does not make sense to maintain anymore.

amp-proxy

Use Amp with your existing Claude Max, ChatGPT Plus, or Gemini subscriptions — no API billing.

Quick Start

git clone https://github.com/johannhipp/amp-proxy && cd amp-proxy
./bin/install.sh
amp-proxy login claude
amp-proxy

Then point Amp at the proxy:

echo '{"amp.url": "http://localhost:18317"}' > ~/.config/amp/settings.json

What It Does

amp-proxy sits between the Amp CLI and the world. It routes LLM calls through your consumer subscriptions (via CLIProxyAPIPlus) while keeping everything else — OAuth, threads, settings, GitHub — on ampcode.com.

Amp CLI
  │
  ▼
amp-proxy (:18317)
  │
  ├── /api/provider/*  ──►  built-in provider gateway   LLM calls (Claude, GPT, Gemini, ...)
  ├── /v1/*, /api/v1/* ──►  built-in provider gateway   LLM streaming
  ├── /auth/*          ──►  ampcode.com                  OAuth (302 redirect)
  ├── /api/internal    ──►  Exa API / stubs              web_search, read_web_page
  └── everything else  ──►  ampcode.com                  threads, settings, GitHub

Single binary. Single port. No vibeproxy needed.

Features

• Built-in provider auth — amp-proxy login claude authenticates with your subscription, no separate app needed
• Multi-provider support — Claude, OpenAI, Gemini, GitHub Copilot, Qwen — with round-robin and failover
• Model remapping — unsupported models (Gemini) get translated to ones you have (Claude, GPT), with full request/response protocol translation across Google GenAI, Anthropic Messages, and OpenAI Chat Completions formats, including streaming
• Configurable — model mappings, providers, API keys, routing strategy — all via optional YAML config
• Web search via Exa — intercepts Amp's web_search and read_web_page tools and routes them through the Exa API
• Credit gate bypass — fakes getUserFreeTierStatus so server-side tools aren't blocked
• Stable tool translation — fixes repeated same-name tool calls with one-to-one ID tracking
• Zero config for common use — works out of the box after login, config file only needed for advanced use

Install

Requires Go 1.21+. Works on macOS and Linux (amd64/arm64).

git clone https://github.com/johannhipp/amp-proxy
cd amp-proxy
./bin/install.sh

This builds amp-proxy and downloads cli-proxy-api-plus (needed for auth flows) into ~/.local/bin.

Override the install directory:

AMP_PROXY_INSTALL_DIR=/usr/local/bin ./bin/install.sh

Auth

Authenticate with one or more providers:

amp-proxy login claude       # Claude Max / Pro
amp-proxy login openai       # ChatGPT Plus / Pro
amp-proxy login gemini       # Gemini
amp-proxy login copilot      # GitHub Copilot
amp-proxy login qwen         # Qwen

Check status:

amp-proxy status

Provider     Status        Account              Expires
────────     ──────        ───────              ───────
claude       ✓ active      user@gmail.com       2025-04-15
openai       ✓ active      user@gmail.com       2025-04-12
gemini       ✗ not authed  -                    -

Tokens are stored in ~/.cli-proxy-api/ — compatible with vibeproxy, so existing tokens carry over.

If amp-proxy was already running when you logged in, restart it — tokens are loaded at startup, so new ones added mid-run aren't picked up until restart.

Usage

# Start with defaults (127.0.0.1:18317)
amp-proxy

# Custom port
amp-proxy serve --port 9000

# With Exa web search
EXA_API_KEY=your-key amp-proxy

# Debug logging
amp-proxy serve --debug

Configuration

Config is optional. amp-proxy works with zero config for most users. An optional YAML file unlocks advanced features.

Auto-detected at ~/.config/amp-proxy/config.yaml (macOS: ~/Library/Application Support/amp-proxy/config.yaml), or set AMP_PROXY_CONFIG.

# Only include what you want to change. Everything has defaults.

# Exa API for web_search tool
exa_api_key: ${EXA_API_KEY}

# Provider routing: "round-robin" (default) or "fill-first"
routing: fill-first

# Custom model remapping
model_remaps:
  - from: gemini-3-flash-preview
    to: claude-sonnet-4-6
    provider: anthropic

  - from: gemini-3-pro
    to: gpt-5.4
    provider: openai

# Direct API keys (skip OAuth)
anthropic_api_keys:
  - key: ${ANTHROPIC_API_KEY}

# Enable/disable providers
providers:
  claude: true
  openai: true
  gemini: false

See config.example.yaml for all options.

Prompt Caching

By default, amp-proxy strips cache_control fields from request bodies before forwarding to providers. This prevents 400 errors on some OAuth routes, but disables Anthropic prompt caching — which can significantly increase token usage in long sessions.

To enable prompt caching, set strip_cache_control: false in your config:

strip_cache_control: false

If you see 400 errors from Anthropic after disabling this, re-enable it.

Model Remapping

When Amp requests a model you don't have (e.g., Gemini), amp-proxy translates the request to a provider you do have. This includes full protocol translation — request body, response body, and streaming — across Google GenAI, Anthropic Messages, and OpenAI Chat Completions formats.

Default mappings (configurable via YAML):

Amp requests	Served by	Provider
gemini-3-flash-preview	claude-sonnet-4-6	Anthropic
gemini-3-flash	claude-sonnet-4-6	Anthropic
gemini-3-pro	gpt-5.4	OpenAI
gemini-3-pro-image	gpt-image-1	OpenAI
anything else unsupported	claude-sonnet-4-6	Anthropic

Web Search

Amp's web_search and read_web_page are server-side tools on ampcode.com, gated by credits. amp-proxy intercepts them and routes through the Exa API instead.

Set EXA_API_KEY to enable. Without it, tools return a stub message.

Endpoints

Path	Method	Description
/healthz	GET	Health check → {"status":"ok"}
/metrics	GET	Request counters (JSON)

CLI Reference

amp-proxy [command] [flags]

Commands:
  serve     Start the proxy server (default)
  login     Authenticate with a provider
  logout    Remove saved auth for a provider
  status    Show auth status for all providers
  version   Print version info

Flags (serve):
  --port <port>      Listen port (default: 18317, env: LISTEN_PORT)
  --addr <addr>      Listen address (default: 127.0.0.1, env: LISTEN_ADDR)
  --config <path>    Config file path (env: AMP_PROXY_CONFIG)
  --ampcode <url>    Ampcode URL (default: https://ampcode.com)
  --debug            Enable debug logging (env: AMP_PROXY_DEBUG)

Development

make test    # run tests
make dev     # run without building
make fmt     # format code
make vet     # run go vet

When running from source (make dev / go run .), amp-proxy login shells out to cli-proxy-api-plus, which the install script normally provides. To use login in dev mode, install the binary yourself:

go install github.com/router-for-me/CLIProxyAPI/v6/cmd/server@latest
ln -sf ~/go/bin/server ~/go/bin/cli-proxy-api-plus

Make sure ~/go/bin is on your PATH. Or just run ./bin/install.sh once to get a released cli-proxy-api-plus into ~/.local/bin.

Migrating from vibeproxy

If you're already using vibeproxy + the old amp-proxy:

1. Install the new amp-proxy: ./bin/install.sh
2. amp-proxy status — your existing tokens in ~/.cli-proxy-api/ are picked up automatically
3. amp-proxy — start the proxy, vibeproxy is no longer needed
4. Stop vibeproxy

Releases

Automated via .github/workflows/release.yml and .goreleaser.yml.

git tag v0.2.0
git push origin v0.2.0