llm-council.dev

What It Is

LLM Council is an open-source developer tool built around a simple idea: instead of having a single large language model answer directly, it sends the same question in parallel to multiple LLMs, such as GPT, Claude, Gemini, and Grok. The models then anonymously review and rank each other’s responses, and a Chairman LLM synthesizes the final answer. It is well suited to code review, architecture decisions, content validation, and complex problem analysis where cross-model verification is useful.

Core Features and Integrations

The tool is available in three forms: MCP Server, HTTP API, and Python Library. The HTTP version provides /health, /v1/council/run, and the SSE streaming endpoint /v1/council/stream, along with Swagger UI, ReDoc, and OpenAPI JSON. Clients can integrate via Python, JavaScript/TypeScript, or cURL. For authentication, it supports a local Bearer token. For model gateways, it supports OpenRouter and Requesty, and can also connect directly to APIs from Anthropic, OpenAI, Google, and others. The n8n integration documentation is fairly detailed, covering scenarios such as Webhooks, HMAC signatures, asynchronous events, ticket routing, and PR review.

Open Source, Self-Hosting, and Pricing

The documentation states that the OSS package is MIT-licensed and that the core algorithm is open source. There is also a proprietary council-cloud for production use cases such as authentication, billing, caching, and audit logs. The OSS version uses BYOK and does not appear to charge by itself, but users need to pay for calls to multiple models. Deployment options include a local HTTP Server, local Docker, Railway, and Render. Its design principles are stateless, single-tenant, no database, and logs output only to stdout.

Pros and Cons

Its strengths are a clear architecture and anonymous peer review, which can reduce model bias. Mechanisms such as Borda ranking, binary arbitration, tie-breakers, and include_dissent make it suitable for engineering decisions. The documentation quality is also strong, with reasonably complete coverage of the API, n8n, security recommendations, and ADRs. The downside is that multi-model deliberation naturally increases cost and latency; the documentation notes that it may take 30–60 seconds. The OSS service also does not provide full production-grade multi-tenancy, persistence, auditing, or billing capabilities.

Who It’s For and Access from China

It is a good fit for AI engineers, architects, DevOps automation teams, and teams that want to introduce multi-model review into Claude Code, n8n, or internal tools. The documentation does not specify accessibility from mainland China. However, because it depends on external APIs such as OpenRouter, OpenAI, Anthropic, and Google, network connectivity and payment availability may be uncertain. If access is restricted, alternatives include building a similar workflow with LangChain, LlamaIndex, or Vercel AI SDK, or connecting to an available local/mainland China model gateway.