claude-code-llm-gateway

Implement POST /v1/messages with stream:true and stream:false support, returning Anthropic message JSON or raw SSE event sequences. Claude Code connects by setting ANTHROPIC_BASE_URL to your gateway, which then routes requests to your upstream model.

Return Content-Type text/event-stream and emit events in order: message_start, content_block_start, content_block_delta, content_block_stop, message_delta, and message_stop. Even if your upstream model is non-streaming, you must synthesize these events.

Yes. When the model needs a tool, return stop_reason tool_use with a tool_use content block; Claude Code sends the result back as a tool_result block in the next user message. The gateway must preserve this conversation structure.

Yes. The gateway is responsible for converting OpenAI or custom upstream responses into Anthropic content blocks and SSE events. Unknown request fields like context_management or speed should be ignored or passed through rather than rejected.

Return Anthropic-style JSON with type error and an error object containing type and message. Use 429 or 529 for retryable rate limiting, 401 for authentication errors, and 5xx for retryable server errors.

P0 endpoints are POST /v1/messages, SSE streaming, and tool round-trips. P1 covers count_tokens and the Files API; P2 is GET /v1/models. OAuth, bootstrap, and session ingress endpoints are explicitly out of scope.