raysonmeng/agent-bridge

AgentBridge – a local bridge that lets Claude Code and OpenAI Codex talk to each other in real time

What it is – AgentBridge is a developer‑side utility that runs a tiny daemon and a command‑line wrapper (abg / agentbridge). It connects two large‑language‑model coding assistants – Claude Code (Anthropic) and Codex (OpenAI) – so they can exchange messages, review each other’s work, and hand‑off tasks without you copying text between terminals.

Why it matters – Normally you would have to run two separate terminals, copy‑paste code, and manually switch when one agent hits its quota. AgentBridge automates the “relay”:

• each agent stays alive as a persistent peer,
• messages are pushed automatically (with a fallback queue),
• a busy‑guard prevents overlapping replies, and
• a built‑in watchdog lets a long job survive subscription‑quota limits by handing the turn to the other agent.

---

Core features (as described in the README)

Feature	What you get
Bidirectional Claude ↔ Codex messaging	Claude receives Codex’s agentMessage payloads as channel notifications; Codex receives Claude’s replies via the reply tool.
Push delivery with fallback queue	If a push fails, the daemon stores the message in memory and retries via get_messages.
Turn coordination	A busy‑guard blocks replies while the opposite side is in an active turn; an inactivity watchdog forces a turn to complete after 5 min of silence.
Multiple pairs side‑by‑side	You can run several Claude+Codex pairs in the same repo; each pair gets its own set of ports (+10 stride from 4500).
Resilient lifecycle	The daemon survives Claude restarts, auto‑reconnects with back‑off, cleans up orphan processes, and offers abg doctor diagnostics.
Thread auto‑resume	abg codex automatically re‑opens the last Codex thread; abg resume prints or runs the needed commands for both sides.
Budget coordination & auto‑resume	When one agent’s subscription quota expires, the bridge pauses that side and lets the other continue, keeping long jobs alive.
Filtering to keep context small	Only agentMessage payloads cross; other tool‑call noise is dropped. A three‑tier marker system ([IMPORTANT], [STATUS], [FYI]) lets you control what is forwarded, buffered, or ignored.
Cross‑network collaboration (preview)	v3 adds a broker‑based “room” so pairs can run on different machines and still exchange messages.

---

Quick‑start workflow (from the README)

1. Install Bun – the runtime required for the daemon.
2. Install the CLI – npm install -g @raysonmeng/agentbridge (post‑install registers a Claude Code plugin).
3. Initialize the project – abg init creates .agentbridge/config.json and checks dependencies.
4. Start Claude – abg claude launches Claude Code with the bridge channel enabled.
5. Start Codex – in another terminal, abg codex launches the Codex TUI connected to the same bridge.

Now you can ask Claude to “propose a task split with Codex for …” and watch the two agents negotiate, implement, and review code without any manual copy‑pasting.

---

How it works under the hood

• bridge.ts – a short‑lived foreground client that Claude loads as a plugin. It forwards Claude’s reply calls to the daemon and receives Codex messages to push back into Claude.
• daemon.ts – a persistent background process that talks to the Codex app‑server (via WebSocket or Unix socket) and holds the authoritative state for the pair.
• Communication flow:
  • Codex → daemon → bridge → Claude (channel notification)
  • Claude → bridge → daemon → Codex (injected turn/start)
• Each message carries a source field so the bridge never echoes a message back to its origin, preventing loops.

---

Installation & prerequisites (as listed)

Dependency	Recommended version	Install command
Bun	v1.3.11+	`curl -fsSL https://bun.sh/install
Claude Code	2.1.269	npm install -g @anthropic-ai/claude-code@2.1.269
Codex CLI	0.154.0	npm install -g @openai/codex@0.154.0

Note: The CLI defaults launch Claude with --dangerously-skip-permissions and Codex with --yolo (i.e., they can run commands and edit files without prompting). Use --safe or set AGENTBRIDGE_SAFE=1 if you need the normal permission prompts.

---

Typical commands

Command	What it does
abg init	Checks deps, registers the Claude plugin, writes .agentbridge/config.json
abg claude [args]	Starts Claude Code with the bridge channel (injects required flags)
abg codex [args]	Starts Codex TUI attached to the bridge; abg codex --new creates a fresh thread
abg resume	Prints or executes the commands needed to re‑attach both sides to the last session
abg pairs	Lists, removes, or prunes multiple Claude+Codex pairs
abg doctor	Read‑only diagnostics (environment, daemon health, logs)
abg budget	Shows each agent’s subscription quota status
abg kill	Gracefully stops the daemon and any managed Codex TUI
abg dev	For developers – registers a local marketplace and forces a plugin sync

---

Who might use it

• Solo developers who already have access to Claude Code and Codex and want them to collaborate automatically.
• Researchers exploring multi‑agent coding workflows, task‑splitting, or cross‑review between different LLM providers.
• Teams that run long‑running code‑generation jobs and need a way to keep work alive when one provider’s quota expires.

---

Limitations (explicitly called out)

• It is local‑only – not a hosted, multi‑tenant service.
• It only supports the two specific agents (Claude Code and Codex); it is not a generic orchestration framework.
• Security boundaries are minimal – the bridge assumes you trust both agents and the workspace.

---

Where to learn more

• Website & demo: https://raysonmeng.github.io/agent-bridge/ (animated replay of a real session)
• Documentation: docs/ folder in the repo – includes a manual, troubleshooting guide, and details on the cross‑network “room” feature.
• Community discussion: LINUX DO developer forum (link in README).

---

TL;DR

AgentBridge is a lightweight, Bun‑based daemon plus CLI that lets Claude Code and OpenAI Codex converse, split tasks, and review each other’s output inside a single development session, handling turn coordination, quota hand‑off, and message filtering automatically. It is aimed at developers who want two LLM coding assistants to work together without manual copy‑pasting.