OpenAI-Compatible API

These endpoints are designed for OpenAI-style client compatibility while routing through cliproxyapi++ provider logic.

Base URL

http://<host>:8317

Authentication

/v1/* routes require a configured client API key:

Authorization: Bearer <api-key-from-config.yaml-api-keys>

Endpoints

POST /v1/chat/completions

Use for chat-style generation.

curl -sS -X POST http://localhost:8317/v1/chat/completions \
  -H "Authorization: Bearer dev-local-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-3-5-sonnet",
    "messages": [{"role": "user", "content": "Give me 3 release notes bullets"}],
    "temperature": 0.2,
    "stream": false
  }'

Example response shape:

{
  "id": "chatcmpl-...",
  "object": "chat.completion",
  "created": 1730000000,
  "model": "claude-3-5-sonnet",
  "choices": [
    {
      "index": 0,
      "message": {"role": "assistant", "content": "..."},
      "finish_reason": "stop"
    }
  ],
  "usage": {"prompt_tokens": 10, "completion_tokens": 42, "total_tokens": 52}
}

POST /v1/completions

Legacy completion-style flow for clients that still use text completion payloads.

POST /v1/responses

Responses-style payload support for compatible clients/workloads.

GET /v1/models

Lists models visible under current configuration and auth context.

curl -sS http://localhost:8317/v1/models \
  -H "Authorization: Bearer dev-local-key" | jq '.data[:10]'

Streaming Guidance

• For SSE, set "stream": true on chat/completions.
• Ensure reverse proxies do not buffer event streams.
• If clients hang, verify ingress/edge idle timeouts.

Claude Compatibility Notes (#145 scope)

• Use canonical OpenAI chat payload shape: messages[].role + messages[].content.
• Avoid mixing /v1/responses payload fields into /v1/chat/completions requests in the same call.
• If you use model aliases for Claude, verify the alias resolves in GET /v1/models before testing chat.
• For conversion debugging, run one non-stream request first, then enable streaming once format parity is confirmed.

Claude OpenAI-Compat Sanity Flow

Use this order to isolate conversion issues quickly:

1. GET /v1/models and confirm the target Claude model ID/alias is present.
2. Send one minimal non-stream chat request.
3. Repeat with stream: true and compare first response chunk + finish reason.
4. If a tool-enabled request fails, retry without tools to separate translation from tool-schema problems.

Minimal non-stream probe:

curl -sS -X POST http://localhost:8317/v1/chat/completions \
  -H "Authorization: Bearer dev-local-key" \
  -H "Content-Type: application/json" \
  -d '{
    "model": "claude-3-5-sonnet",
    "messages": [{"role":"user","content":"reply with ok"}],
    "stream": false
  }' | jq

Common Failure Modes

• 401: missing/invalid client API key.
• 404: wrong path (use /v1/... exactly).
• 429: upstream provider throttling; add backoff and provider capacity.
• 400 model_not_found: alias/prefix/config mismatch.
• 400 with schema/field errors: payload shape mismatch between OpenAI chat format and provider-specific fields.

Related Docs

• Provider Usage
• Routing and Models Reference
• Troubleshooting