deliberation

deliberation-mcp

Get a second opinion from GPT, Gemini, Grok, and 400+ OpenRouter models - and let them reach a cross-model consensus - from inside any MCP host (Claude Code, Cursor, Codex, Kiro, VS Code, Windsurf, Zed, ...).

One model is a guess. Three that agree is a plan.

This is the standalone MCP server. For the Claude Code plugin, one-click install buttons, and full docs, see the repo: https://github.com/antonbabenko/deliberation

What it does

You stay the primary agent. When a task needs a second opinion or cross-model review, call one of the tools below, read the result, and apply your own judgment. Eight expert personas (architect, code reviewer, security analyst, and five more) shape the review. GPT and Gemini can also implement changes; Grok and OpenRouter only advise.

Install

Add the server to your MCP host (most hosts use the mcpServers key):

{
  "mcpServers": {
    "deliberation": {
      "command": "npx",
      "args": ["-y", "@antonbabenko/deliberation-mcp"],
      "env": {
        "XAI_API_KEY": "xai-...",
        "OPENROUTER_API_KEY": "sk-or-v1-..."
      }
    }
  }
}

Per-host config location and key:

Provider setup

Set only the providers you use:

• GPT - install the Codex CLI and run codex login (codex login --device-auth on a remote machine), or set CODEX_ACCESS_TOKEN (ChatGPT Business/Enterprise). OPENAI_API_KEY is never used.
• Gemini - install the Antigravity CLI (agy) and run it once to sign in.
• Grok - set XAI_API_KEY (https://console.x.ai).
• OpenRouter - set OPENROUTER_API_KEY and declare models in ~/.config/deliberation/config.json (Windows: %APPDATA%\deliberation\config.json; override with DELIBERATION_CONFIG). OpenRouter is advisory-only.

A starter config writer ships as a bin:

npx -y --package @antonbabenko/deliberation-mcp deliberation-setup

It writes a starter config.json (never overwrites an existing one).

Tools

• ask-all - one question to every configured provider in parallel; get each answer back independently (no cross-talk).
• consensus - run the full multi-round convergence loop with a provider arbiter (blind pass + peer review -> adjudicate -> revise) and return the converged verdict. Pass synthesizeAlways: true for a single synthesis pass instead of the loop (best for open questions), or maxRounds to cap the loop.
• consensus-step - drive the loop yourself as the arbiter, one action per call (init -> record_blind -> dispatch_peers -> submit_adjudication -> submit_revision).
• ask-gpt / ask-gemini / ask-grok / ask-openrouter - one question to one provider.
• Experts: architect, plan-reviewer, scope-analyst, code-reviewer, security-analyst, researcher, debugger, writing-editor - call directly, or pass expert on the fan-out tools to apply one persona to every delegate.
• Session tools (opt-in sessions.persist): session-get, session-revisit, session-annotate.

Every tool takes a prompt - give it full context (the goal, relevant code/paths, prior attempts), since the experts do not share your session.

How consensus stays honest

The orchestrator commits a blind verdict before seeing the panel, cannot reach consensus on its own vote alone (at least one external must approve), and records a reason for every dismissed issue. The loop stops on agreement or at consensus.maxRounds (default 5).

More

Configuration reference, the Claude Code plugin, per-host guides, and architecture docs are in the repo: https://github.com/antonbabenko/deliberation (see docs/tool-guide.md for the tool guide and TECHNICAL.md for the full config reference).

License

MIT