Connecting tools

Any client that lets you set an OpenAI base URL works:

Setting	Value
Base URL	http://localhost:3456/v1
API key	the value of auth.llmux_key in your config
Model	anything; llmux replaces it with the model it selects

The model field is only honoured when it names an alias from the aliases section; otherwise selection is dynamic. To pin a model, use x-llmux-model.

Request headers

Header	Effect
x-llmux-tool	Names the calling tool in the log (unknown if absent).
x-llmux-project	Applies that project’s scope from projects.<name> and tags the log entry.
x-llmux-profile	interactive, balanced or batch; overrides routing.default_profile.
x-llmux-session	Keeps an agent loop on the same model while the budget allows.
x-llmux-model	Forces a catalog model (model or provider/model) or an alias.
x-llmux-no-cache	true, 1 or yes: skip cache lookup and store.
x-llmux-no-fallback	true, 1 or yes: try only the primary model.
x-llmux-max-cost	Reject with 402 if the estimated cost in USD exceeds this value, e.g. 0.05.

A forced model does not bypass hard constraints. It is still checked for provider and key readiness, tool support, context fit, privacy local_only and remaining budget; a violation is rejected and logged with forced: true, result: rejected.

Response headers

Header	Meaning
x-llmux-cache	hit when the response came from the cache.
x-llmux-echo	1 when the built-in echo provider answered.

Agent loops and tools

Requests with tools, tool_choice, functions or tool messages in the history are routed only to models with the tools capability (supports_tools: true). The tool schema is passed through unchanged and counted in the token estimate, so models whose context window is too small drop out as the history grows. Send x-llmux-session: <id> to keep tool calls and tool results on one model.

Example

curl -s http://localhost:3456/v1/chat/completions \
  -H 'Authorization: Bearer mp_dev_changeme' \
  -H 'content-type: application/json' \
  -H 'x-llmux-tool: curl' \
  -H 'x-llmux-model: echo-ultra' \
  -d '{"messages":[{"role":"user","content":"Hello"}]}'

The forced model example shows the response.