agent-bridge / site

# AgentBridge

> A local, MIT-licensed developer tool that lets two AI coding agents — Claude Code (Anthropic) and Codex (OpenAI) — collaborate in real time inside one working session. It relays messages between an MCP channel and the Codex app-server protocol so the two agents can cross-review each other's diffs, split tasks, and hand work off across quota limits. It runs entirely on the developer's machine, has zero runtime dependencies, and is not a hosted service.

## Facts

- Name: AgentBridge
- License: MIT
- Latest version: v0.1.24
- Install (CLI): `npm install -g @raysonmeng/agentbridge`
- npm package: @raysonmeng/agentbridge
- Repository: https://github.com/raysonmeng/agent-bridge
- Website: https://raysonmeng.github.io/agent-bridge/
- Author: raysonmeng (https://x.com/raysonmeng)
- Platforms: macOS and Linux (Windows not officially supported yet)
- Runtime: Bun (required — Node.js alone is not sufficient)
- Tests: 1,700+
- Runs 100% locally; no accounts, API keys, or servers needed for AgentBridge itself (uses your existing Claude Code and Codex logins). Raw work never leaves the machine.

## What it is

- A local bridge that connects Claude Code and Codex in one workflow, in the same working session.
- Two-process architecture: a foreground MCP client (`bridge.ts`, started by Claude Code via the AgentBridge plugin) and a persistent background daemon (`daemon.ts`) that owns the Codex app-server proxy and bridge state. The daemon survives Claude Code restarts; the foreground reconnects with exponential backoff.
- An experimental setup for human-in-the-loop collaboration between multiple agents. The human steers; the agents coordinate.

## What it is NOT

- Not a hosted service or multi-tenant system.
- Not a generic orchestration framework for arbitrary agent backends.
- Not a hardened security boundary between tools you do not trust.

## Three core benefits

- Cross-review, not just cross-talk: Codex implements; Claude reviews the diff in real time inside the same session and pushes change requests straight back into Codex's thread. Two providers holding each other accountable, no copy-paste.
- Task splits from one prompt: ask either agent to propose a division of labor with the other; they negotiate who does what before writing code.
- Quota relay for overnight runs: when one side's subscription window runs dry it stops cleanly at a turn boundary, writes a checkpoint, and hands off, so a long job keeps moving instead of dying at a limit. Requires the companion tool agent-quota-guard.

## How it compares

- vs two terminals + copy-paste: with two terminals the human is the message bus, ferrying text by hand. AgentBridge automates the relay — messages flow on their own, a busy-guard blocks replies during an active turn, and noisy intermediate events are filtered.
- vs a one-way delegation plugin (e.g. openai/codex-plugin-cc): those let a host call Codex and get one answer back (request in, response out, no standing peer). AgentBridge keeps both agents live as persistent peers, and either side can push a message mid-turn, not only at call boundaries.
- vs an external orchestrator: a god-process scheduling dumb terminals is top-down (one brain, N workers that never talk). AgentBridge is peer-to-peer — two full agents converse in-session, propose their own splits, and review each other.

## Quick start (4 steps)

1. Install Bun: `curl -fsSL https://bun.sh/install | bash`
2. Install the CLI: `npm install -g @raysonmeng/agentbridge` (postinstall best-effort registers the Claude Code plugin)
3. Start Claude Code with the bridge channel: `abg claude`
4. In another terminal, start Codex on the same bridge: `abg codex`

`abg` is a short alias for `agentbridge`. Both agents launch with max-permission defaults so an unattended pair does not stop for each prompt; add `--safe` (or set AGENTBRIDGE_SAFE=1) for normal permission prompts. Only run the default mode in a workspace you trust.

## Key CLI commands

- `abg init` — install plugin, check dependencies, generate `.agentbridge/config.json`
- `abg claude [args]` — start Claude Code with the push channel enabled
- `abg codex [args]` — start Codex TUI connected to the AgentBridge daemon (auto-resumes the pair's last thread; `--new` for a fresh one)
- `abg budget [--json]` — both agents' subscription quota snapshot (5h/weekly windows, drift, pause state)
- `abg doctor [--json]` — read-only diagnosis of env, daemon health, build drift
- `abg pairs` — list/remove/prune registered pairs (one pair per project directory)
- `abg kill [--all]` — stop the pair's daemon and managed Codex TUI

## How it was built

AgentBridge was largely built by Claude Code and Codex collaborating through it — the very tool they were building together. Every PR written by one agent was reviewed by the other, with a human steering. It is its own proof of concept.

## Roadmap

- More adapters: which agent is wired in next is a public vote at https://github.com/raysonmeng/agent-bridge/issues/212 (candidates include OpenCode and Gemini CLI).
- Capability mesh: rooms of agents across machines that advertise capabilities and route work accordingly (v3 cross-network preview ships behind experimental commands today).

## Companion tool

- agent-quota-guard (https://github.com/raysonmeng/agent-quota-guard) — enables the quota-relay / slowdown-line / auto-resume feature. Optional; everything else works without it.