JoaoBerne/cli-bridge-mcp

cli-bridge

A Model Context Protocol server that lets the AI assistant you're talking to consult the other AI CLIs installed on your machine — Claude Code, Codex, Gemini, Mistral, opencode, Ollama, Apple fm, … Each lane spawns the official CLI as a subprocess: no API keys, no token extraction, read-only by default.

Dependencies: the Python stdlib and mcp (1.x or 2.x). No daemon.

What it does

ask_gemini(task="find the bug across ./src", cwd="path/to/repo")     # one lane (1M-token context)
ask_apple(task="what's wrong in this UI?", images=["shot.png"])      # vision, on-device, $0
ask_all(task="…")                                                    # every free lane in parallel + disagreement score
ask_cascade(task="…")                                                # cheapest→strongest, skips cooled-down lanes
ask_best(task="…", mode="deep")                                      # router picks; rate_lane teaches it
ask_build(lane="opencode", task="add retry with backoff")            # build in a throwaway worktree → diff
review_diff(base="origin/main", focus="security")                    # multi-model review, severity-ranked
debate(task="which migration strategy?", vote="borda")               # N blind answers, peer-ranked
workflow(preset="converge", task="is this migration safe?")          # author → blind arbiter → cross-family peers
git_text(kind="commit")                                              # Conventional Commit from the staged diff

Every ask<lane returns a thread id; reuse it (even on another lane) for a multi-model conversation that survives /compact. conversations lists or replays threads.

Tools

15 fixed tools + one ask<lane per installed CLI:

• Consult: ask<lane, askall, askcascade, askbest, conversations, listmodels
• Build: askbuild (mode=isolated → diff, mode=direct → zone-guarded writes; with mode=direct, async=true → steerable via job)
• Review: reviewdiff (focus=code|security), debate (vote=judge|borda), workflow (presets: refineplan, mapreview, researchverify, fanoutcompare, converge, premortem, testplan, challenge), gittext (kind=commit|pr)
• Operate: job (action=status|result|cancel|list|tail|steer), ratelane, setlanecost, doctor, setup

CLIBRIDGETOOLS=all adds batchrun (journaled fan-out) and resetlanestate; CLIBRIDGETOOLS=default,batchrun extends the default; a plain comma list is exactly those. Usage and lane health are MCP resources (cli-bridge://usage-summary, cli-bridge://lane-stats). Full reference: docs/TOOLS.md.

There is also a human CLI, cli-bridge doctor|ask|ask-all|ask-best|build|review-diff|security-review|test-plan|premortem|stats|usage|jobs|set-cost (--json where it makes sense). cli-bridge build <lane "<task" prints the worktree diff; --apply lands it as unstaged changes.

Writing code safely

• isolated (default): the delegate edits a throwaway git worktree and you get a diff. Your tree is untouched.
• direct: writes real files, but only inside the zone you declare, behind a per-zone lock with a post-turn zone-violation check; undo is zone-scoped. async=true makes it steerable (job action=tail|steer) with an executable Definition-of-Done (dodcmd).

CLIBRIDGEVERIFYPLANREADONLY=1 flags (never reverts) a read-only delegate that wrote files anyway. Re-entry is depth-capped (CLIBRIDGEMAXDEPTH, default 1). Delegates run in the caller's cwd, else CLIBRIDGEDEFAULTCWD, else the host's MCP workspace root.

Install

Prerequisites: Python 3.12+, uv, and at least one AI CLI installed and logged in.

uvx --from cli-bridge-mcp cli-bridge doctor        # what cli-bridge can see (--deep probes each lane)

Wire it into your host: