Default base URL:
http://127.0.0.1:11435
Start a server:
tapioca serve MODEL --host 127.0.0.1 --port 11435Local compatibility clients may send any non-empty API key such as
tapioca-local. Tapioca does not require a secret for a loopback-only server.
GET /healthWait for a successful response before sending inference requests. A TCP connection alone does not mean the model has finished loading.
GET /v1/modelsUse this endpoint to confirm the active served model. Refresh with
tapioca catalog update, then use tapioca catalog to discover models that
can be installed.
POST /v1/chat/completions
Content-Type: application/jsonSupports messages, streaming, tools, tool_choice, assistant tool calls,
tool-result messages, and reasoning-capable model output.
See chat-completions.json and tool-calling.json.
POST /v1/responses
Content-Type: application/jsonUse input for a string or Responses-style input items. Tapioca translates
function tools and tool results to the active model server.
See responses-api.json.
POST /v1/messages
Content-Type: application/jsonThis compatibility route supports Claude Code launch flows and Anthropic-style messages and tool-use blocks.
- Send the available function definitions with the user request.
- If the assistant returns tool calls, validate each function name and JSON argument object.
- Execute only tools allowed by the host agent's permission policy.
- Return a tool-result message with the matching call ID.
- Repeat until the model returns a final answer or the host's iteration limit is reached.
Tapioca transports tool calls; it does not execute them.