cli-bridge — your AI assistant, but it can phone a friend

// what it is

An MCP server that orchestrates the AI CLIs you've already installed and logged into — Claude Code, Codex, Gemini CLI, opencode — from whatever assistant you're talking to. It never holds an API key and never extracts a token: it spawns each vendor's official CLI, exactly as you would by hand. Reviews fan out to a role-diverse council in parallel and come back merged, deduped and severity-ranked.

// 30 seconds, $0

security-review demo: a committed auth bypass caught independently by two models

Real run (2.5× speed): a committed auth bypass — two free models flag it blocker independently; usage shows the receipts.

// install

# zero-install run
uvx cli-bridge-mcp

# wire it into Claude Code
claude mcp add cli-bridge -- uvx cli-bridge-mcp

# or any MCP host — codex, gemini, opencode…

// why this one

Ban-safe by design

Spawns official CLIs. No OAuth-token extraction, no key reuse, nothing that gets accounts flagged.

Learning router

Rate answers 1–5; ask_best learns which model wins each task-type on your machine. Local signal, survives restarts.

Cost governance

Sourced per-plan tiers, free-first councils, a hard daily cap and local receipts — spend is bounded by design.

Safe writes

Build agents run in a throwaway git worktree and hand back a diff. Your live repo is never touched.

Memory that survives

Round-table threads persist in sqlite — across host context resets and restarts.

Measured, not asserted

cli-bridge eval tests the council-vs-single claim deterministically and publishes the result either way.

// the honest part

"More models = better" is fragile — big models share training data, so their errors correlate. We measured our own claim: a diverse council did not catch more bugs than one strong model — it cut false alarms ~2×. The numbers are in BENCHMARKS.md, and the harness ships so you can run it on your own CLIs.