A single OpenAI-compatible endpoint that bridges Cursor, Claude Code, Gemini CLI, Codex, and OpenRouter. NestJS + Fastify on port 3900, LIVE on connector.arcanada.ai. Public repo under MIT. An internal benchmark documented the hard tradeoff: CLI connectors lack json_schema and concurrency, so they cannot replace OpenAI API for frameworks like Graphiti, Cognee, or LangChain agents — use OpenRouter or direct API for structured-output workloads.
Capabilities
- OpenAI-compatible /chat/completions endpoint (HTTP 201 on /execute)
- Connectors: Claude CLI, Cursor, Gemini, Codex, OpenRouter
- Per-tenant rate limits and quotas
- Token accounting per request and per tenant
- Public MIT-licensed repo (Arcanada-one/model-connector)
- 104 tests covering all four CLI connectors
- Documented capability matrix (json_schema / concurrency support per connector)
Current autonomy level
Weakest link
CLI connectors are inherently brittle — agents may exit 0 on logical error, structured output is not supported, concurrency is 1-at-a-time. No automatic fallback chain between providers; if Claude CLI hangs, the request hangs.
Roadmap to L3
- L3 lift — connector self-test on startup, schema-validated outputs via Zod, structured failure classification per connector.
- L4 lift — provider fallback chain (Claude → OpenRouter → OpenAI) with circuit breaker per connector; hard cost CB per tenant.
- L5 candidate — automatic provider routing based on task complexity, budget, and structured-output requirement.