English · 繁體中文

claude-bridges

MCP servers that let Claude Code delegate to other coding CLIs — cross-model second opinions, structured reviews, and a multi-engine consultation panel.

Most people use AI to autocomplete or to ask one model one question. I use multiple models as a structured thinking system with checks and balances: debate, execute, review, cross-check, iterate.

At a glance

Related MCP server: personal-mcp

How it works

flowchart TD
    A["🧠 I think / identify problem"] --> B["Claude Opus 4.6 @ claude.ai<br/>(debates with me, holds personal memory)"]
    B --> C["Claude Code: Fable refines plan<br/>(optional, costly)"]
    C --> D["Claude Code: Sonnet 5 executes<br/>(Opus 4.8 advisor) → HANDOFF.md"]
    D --> E["Opus 4.6 reviews with me"]
    E --> F{"Cross-model review<br/>(consult-bridge)"}
    F -.-> G["Gemini (agy-bridge)<br/>(default while quota lasts)"]
    F --> H["GitHub Copilot CLI"]
    F --> I["OpenAI Codex CLI"]
    F -.-> J["Grok / others<br/>(planned, not yet wired)"]
    G & H & I --> K["Synthesize + next iteration"]
    K --> A

    style A fill:#2d3748,color:#fff
    style F fill:#4a5568,color:#fff
    style J fill:#718096,color:#a0aec0,stroke-dasharray: 5 5

This isn't "ask ChatGPT." Each model has a role, a sandbox, and a review gate.

The panel is the engines in CONSULT_ORDER (default copilot,codex,deepseek,agy,nim,openrouter,opencode,azure,gemini). In mode: all (the default) one prompt goes to all of them in parallel and every successful answer comes back labelled by engine. In mode: first they run as a sequential fallback chain and the first success wins. A failed engine fails over to the next, and the trail is appended to the result.

deepseek, nim, openrouter and dsapi get no cwd, shell or file access — they only ever see the prompt text, so their prompts must be self-contained.

The full engine table, each fallback chain, and the per-engine timeouts are in docs/engines.md.

The engineering story

A single answer hides its own blind spots. The panel exists so that no single model's output ships unchecked.

A 2026-10-27 eval used a code-review prompt with planted bugs to pick the NIM chain. Ultra answered in 27-79 s, minimax-m3.1-flash in 93 s and super-vl in 57 s, and all three found the bugs. glm-5.3 (396 s), glm-5.3-flash (319 s) and step-5-preview (228 s) were too slow or too erratic for the first steps; openai/gpt-oss-20b answered in 5-10 s and closes the chain. The same day, the OpenRouter free models answered a small code-review prompt in 2-21 s.

The cause of that spread is the free tier itself: models stall, vanish (HTTP 404, or 410 after end of life) or return reasoning only, and latency varies a lot run to run because the free queue is shared.

The fix is a fallback chain per free-tier engine, with a per-attempt timeout and a rule that a bare nim or openrouter may land on a different model. A test enforces that the OpenRouter and NIM chains share no model family, so the panel never counts one voice twice. A quota limit that names its own reset time becomes a lock, so the next run skips that engine until then.

What stays open: Grok and others are in the loop diagram but have no bridge yet, and agy leaves the default order once its Google AI Plus quota is used up.

The wider division of labour is in docs/why.md: Opus 4.6 holds the plan and the memory, Sonnet 5 writes the code with Opus 4.8 as advisor, and consult cross-checks before anything lands. Each delegation also carries a role prompt.

Build / run

npm install
node consult-cli.js --mode all --order copilot,codex "prompt"
npm test

One npm install at the root serves all servers — Node resolves node_modules by walking up from each index.js.

• MCP registration, the CLI logins, and the "timeout": 660000 that nim needs: docs/setup.md.

• All flags, round 2, quota locks and debate mode: docs/consult-cli.md.

• Environment variables and their defaults: docs/configuration.md.

• consult-cli.js --json prints one {ok,code,message,data} object; the contract is in docs/cli-json.md.

Repo layout

claude-bridges/
  package.json            shared deps (@modelcontextprotocol/sdk, zod)
  node_modules/           one install serves all bridges
  lib/
    common.js             shared MCP boilerplate, CLI runner, cross-platform exe resolution
    engines.js            engine runners shared by consult-bridge and consult-cli
    args.js               consult-cli argument parsing
    updates.js            throttled update check + --update
  consult-cli.js          same panel as a plain command (no MCP schema cost)
  test/                   node --test suite (npm test)
  copilot-bridge/
    index.js              standalone MCP server → copilot_exec tool
  codex-bridge/
    index.js              standalone MCP server → codex_exec tool
  consult-bridge/
    index.js              aggregator MCP server → consult tool
  agy-bridge-vendored/    vendored copy of the agy-bridge (Gemini)

There are three MCP servers. consult-bridge is the one in active use; copilot-bridge and codex-bridge are legacy (kept working, not registered by default). Tool names, default postures and binary resolution are in docs/bridges.md.

Limits and open items

• Grok and others are in the aspirational loop but do not have bridges yet.

• agy is in the default order only while its Google AI Plus quota lasts (since 2026-10-27).

• opencode free models may be logged or used by the provider: send non-private prompts only.

• NIM free-trial use is subject to the NVIDIA API Trial Terms (evaluation, non-production use), and the budget is per key, so any other application using the same key shares it.

• NIM's /v1/models also lists models the key cannot call, and a model stuck in the free-tier queue still fails at about 300 s.

• The dsapi 300 requests/day warning is a soft alarm only: nothing is blocked.

License

MIT License. See LICENSE.