raysonmeng/agent-bridge

What it solves

AgentBridge enables bidirectional, real-time communication between Claude Code and Codex within a single working session. It eliminates the need for manual copy-pasting between two different AI coding agents, allowing them to collaborate, review each other's work, and coordinate tasks automatically.

How it works

The project uses a two-process local bridge consisting of a foreground MCP client (bridge.ts) and a persistent background daemon (daemon.ts). The daemon intercepts Codex output and pushes it to Claude as channel notifications, while Claude's replies (via a reply MCP tool) are injected back into the Codex thread. To prevent context bloat, the bridge only forwards meaningful agentMessage payloads and uses a three-tier marker routing system ([IMPORTANT], [STATUS], [FYI]) to filter noise.

Who it’s for

It is designed for developers who use both Claude Code and Codex and want to establish a peer-to-peer collaboration workflow where agents can cross-review code or hand off tasks based on subscription quota limits.

Highlights

• Bidirectional Messaging: Real-time communication between Claude Code and Codex without human relay.
• Cross-Review Workflow: One agent can implement a feature while the other reviews the diff and pushes change requests back.
• Context Management: Filters out tool-call noise and reasoning deltas to keep context windows lean.
• Quota Relay: Integrates with agent-quota-guard to hand off tasks between agents when one hits a subscription limit.
• Turn Coordination: Includes a busy-guard and inactivity watchdog to prevent message collisions and deadlocks.