Anyone can call a second model. What is hard is carrying the review findings from one model into another model's pre-commit check without re-explaining everything, and that is what continuation_id does here: a codereview, a planner run and a precommit pass stay in one thread across different tools and different providers. Your agent stays in charge and does the actual work — the other models give perspectives on subtasks. Two things to size up first: each consult costs tokens on your own key, and with a native key and an OpenRouter key both configured, the native provider wins when a model name appears in both.
A Python server that gives your coding agent structured workflows and the ability to consult other models inside them, with one conversation thread shared across every tool and model involved.
- codereview and precommit — a staged walk through the code, then validation of the changes before they land
- debug — an expert debugging pass that takes the error, the context and what you already tried
- planner — a complex project broken into ordered steps, revisable as you go
- consensus — the same question put to several models and their positions gathered
- analyze — file and codebase analysis, with a focus you can set
- refactor — restructuring with decomposition as the priority
- testgen — tests generated with the edge cases spelled out
- secaudit — a security pass with OWASP framing, threat level and compliance requirements as parameters
- docgen — documentation generated with complexity and control flow accounted for
- tracer — call-flow mapping and dependency tracing
- thinkdeep and chat — extended reasoning and ordinary collaborative thinking
- challenge — a prompt that pushes back instead of agreeing with you
- A thinking_mode dial and use_websearch on the tools that benefit, plus a continuation_id that carries the thread between them
At least one model API key of your own — OpenRouter for several providers behind one key, or native keys for Gemini, OpenAI, X.AI or DIAL. A custom endpoint works too, so Ollama, vLLM, LM Studio or anything OpenAI-compatible can serve the models locally. Python 3.10 or higher, with uv for the quick install; Windows needs WSL2 for the Claude Code CLI path. Every consult bills to your own provider account.
One command plus a key — exec $(which uvx || echo uvx) --from git+https://github.com/BeehiveInnovations/zen-mcp-server.git zen-mcp-server, then supply credentials