Skip to main content
StateSet Agents is a Python framework, and most of it is driven from Python or the CLI. It also ships a FastAPI gateway, and this tab is that gateway’s contract: create agents, chat with them, start and watch training jobs, and serve a trained model behind endpoints that Anthropic and OpenAI clients already know how to call.
The server refuses to start with authentication required and no credential source configured. Set API_KEYS (comma-separated) or API_JWT_SECRET, or API_REQUIRE_AUTH=false for local work. Rate limits default to 60 requests a minute per key and 30 unauthenticated, reported in X-RateLimit-Limit, X-RateLimit-Remaining and X-RateLimit-Reset.

​
What is here

Agents

Create, list, inspect and delete agent configurations.

Conversations

Multi-turn chat with an agent, with the transcript kept server-side.

Training

Start a GRPO or other RL training job, poll its status, cancel it.

Messages (Anthropic-compatible)

POST /v1/messages against the configured inference backend; OpenAI-style input accepted.

Chat completions (OpenAI-compatible)

POST /v1/chat/completions and GET /v1/models — point an OpenAI SDK at the server.

Observability

Health, readiness and liveness probes, Prometheus metrics, circuit-breaker state.

​
Serving a trained model

The /v1/messages and /v1/chat/completions routes are thin: they forward to whatever INFERENCE_BACKEND points at — a vLLM server carrying the model you trained — and translate the request and reply into the shape the caller expects.
Add "stream": true to either endpoint for server-sent chunks; set INFERENCE_STREAM_INCLUDE_USAGE=true to have the backend report token usage in the stream when it supports it.

​
Errors and versioning

Every error is one envelope:
GET /api/docs/errors lists the codes a running server can return, and GET /api/docs/changelog what changed between its versions. Deprecated routes answer with a Deprecation header carrying the sunset date.

​
Not in this reference

  • The Training Lab router (/api/lab/*) — the backend of the separately published dashboard and mobile apps. It is behind API_ENABLE_TRAINING_LAB, off by default, and has no deployment path today.
  • The MCP server and CLI, which wrap the framework rather than this gateway — see MCP server and CLI.
The server’s own OpenAPI leaves twenty-two responses untyped. Nineteen of them are typed here from the handler that produces them (spec/overlays/agents.json in the docs repository); the three that remain — the two DELETEs and GET /v1/models — say so on their pages.