open-responses-server CLAUDE.md

CLAUDE.md

This file provides guidance to Claude Code (claude.ai/code) when working with code in this repository.

What This Project Is

Open Responses Server is a proxy that translates OpenAI's Responses API into Chat Completions API calls, allowing any OpenAI-compatible backend (Ollama, vLLM, LiteLLM, Groq, etc.) to serve the Responses API. This enables tools like OpenAI's Codex CLI to work with self-hosted or third-party LLMs.

Build & Run

# Install from source (uses uv)
pip install uv
uv venv
uv pip install -e ".[dev]"

# Start the server
otc start
# Or directly:
uv run src/open_responses_server/cli.py start

Testing

# Run all tests (creates venv, installs deps, runs pytest + verification scripts)
./run_tests.sh

# Run tests manually (after activating venv)
python -m pytest tests/ -v

# Run a single test file
python -m pytest tests/test_server.py -v

# Run a specific test
python -m pytest tests/test_cli.py::TestCLI::test_start_server_imports -v

# Coverage
python -m pytest --cov=open_responses_server tests/

Linting

# The project uses flake8 + bandit + pylint
flake8 src/
pylint src/open_responses_server/
bandit -r src/

Architecture

The server is a FastAPI application with two main API paths that both proxy to a backend LLM:

Request flow:

Client (Codex, etc.)
  |
  v
api_controller.py  -- FastAPI app with route definitions, CORS, startup/shutdown
  |
  +-- POST /responses  -->  responses_service.py
  |     Converts Responses API format to Chat Completions format,
  |     streams SSE events back in Responses API format
  |
  +-- POST /v1/chat/completions  -->  chat_completions_service.py
  |     Proxies with MCP tool injection and automatic tool-call loops
  |
  +-- GET/POST /{path}  -->  Generic proxy to backend (e.g. /v1/models)

Key modules:

• api_controller.py - Route definitions. server_entrypoint.py is the uvicorn entry point that imports from api_controller.
• responses_service.py - Converts Responses API requests to Chat Completions format (convert_responses_to_chat_completions), processes streaming Chat Completions responses back into Responses API SSE events (process_chat_completions_stream). Maintains in-memory conversation_history keyed by previous_response_id.
• chat_completions_service.py - Handles /v1/chat/completions with MCP tool injection. Implements a tool-call loop (up to MAX_TOOL_CALL_ITERATIONS) for both streaming and non-streaming modes.
• common/mcp_manager.py - MCPManager singleton manages MCP server lifecycle (stdio, sse, streamable-http transports), tool discovery/caching with periodic refresh, and tool execution. MCPServer wraps individual server sessions.
• common/llm_client.py - LLMClient singleton wrapping httpx.AsyncClient, pointed at OPENAI_BASE_URL_INTERNAL.
• common/config.py - All configuration via environment variables (loaded from .env via python-dotenv). Key vars: OPENAI_BASE_URL_INTERNAL, OPENAI_API_KEY, MCP_SERVERS_CONFIG_PATH, MAX_TOOL_CALL_ITERATIONS.
• models/responses_models.py - Pydantic models for Responses API request/response/streaming types.