AgentBridge: a bridge between Claude Code and Codex
AgentBridge
A local bridge linking Claude Code and Codex in one session so the agents message each other and review each other's work
- 416 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The launch commands switch off permission prompts for both agents by default, so they run commands and edit files without your confirmation
- A background process listens on local ports from 4500 and relays messages between the agents
- One agent's message content flows into the other agent's thread, so instructions inside it may be acted on
Install
- Works in
In your terminal, with SkillFoxx CLI
npx skillfoxx add plugins/agentbridgeDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The launch commands switch off permission prompts for both agents by default, so they run commands and edit files without your confirmation
- A background process listens on local ports from 4500 and relays messages between the agents
- One agent's message content flows into the other agent's thread, so instructions inside it may be acted on
This is third-party code. Review the repository files before installing.
What it does
AgentBridge keeps Claude Code and Codex in contact inside one working session. A background process intercepts Codex output and pushes it into the Claude Code session as channel notifications. Claude replies through the reply MCP tool and the bridge injects the answer into the Codex thread. If a notification fails, the message waits in a queue that the get_messages tool drains. The developers describe three scenarios. In the first, Codex writes the code, Claude reviews the diff in the same session and pushes change requests back into the Codex thread. In the second, you ask one agent to propose a division of labor and the two agree who does what before any code is written. In the third, when one side runs out of its subscription window it stops at a turn boundary and hands the task to the other, so an overnight run does not die at a limit. Only final messages cross between the agents. Tool calls and reasoning stay with each side, and each keeps its own context window. In the default mode the [IMPORTANT] marker is forwarded at once, [STATUS] is batched into a summary and [FYI] is dropped. The AGENTBRIDGE_FILTER_MODE=full variable switches to the unfiltered stream.
Who it is for. For developers who pay for both Claude Code and Codex and want the two agents to check each other's decisions without copying text between terminals by hand.
Good fit when
- Codex implements a task and Claude should review the changes in the same session and push remarks straight into the Codex thread
- You want to split a large task between two agents and let them settle the distribution themselves
- A long overnight run where, once one agent exhausts its limit window, the other picks the task up
- You need a standing pair of agents in one project that survives a Claude Code restart
- Several projects run in parallel and each folder needs its own Claude and Codex pair
Not a fit when
- You need a hosted service for several users: the bridge is local and its authors describe it that way
- You need a general orchestrator for arbitrary agents: the tool links only Claude Code and Codex
- The working folder holds code you do not trust: by default the agent pair runs commands without prompts
- You have a subscription to only one of the two agents
- You cannot install the Bun runtime on the machine: without it the bridge's service processes do not start
Example request
Propose a task split for this module to Codex, then let Codex write its part while you review the changes.Limitations
Both agents are required, each with its own account, plus the Bun runtime (1.3.11 or newer is recommended): Node.js alone is not enough. The authors' recommended versions at the time of checking are Claude Code 2.1.269 and Codex CLI 0.154.0. The OpenAI and Anthropic APIs are unavailable from Russia, so the bridge fits only where both agents already work.
The authors call the project an experimental local tool. It is not a hosted service and does not serve as a security boundary between tools you do not trust. The notification channel in Claude Code belongs to a research preview, and running from source adds a flag that loads development channels. Cross-machine messaging through a broker is marked as a preview in version 0.1.31.
How to disable. Stop a pair with abg kill, or every pair at once with abg kill --all. Remove the plugin from Claude Code through the /plugin menu and the package with npm uninstall -g @raysonmeng/agentbridge. The .agentbridge folder in the project can be deleted by hand.
FAQ
How do I connect Codex to Claude Code with AgentBridge?
Install Bun, then npm install -g @raysonmeng/agentbridge: the install registers the plugin in Claude Code itself. In the project folder run abg init, then abg claude in one terminal and abg codex in another. The bridge starts the daemon on its own and reconnects after a restart.
How is AgentBridge different from a plugin that calls Codex from Claude Code?
A one-way plugin sends a request and gets one answer back, with no standing peer on the other side. AgentBridge keeps both agents live, and either can send a message in the middle of the other's turn, such as a code remark while the other is still working.
Does the context balloon when the agents message each other constantly?
The bridge passes messages, not context. Each agent keeps its own window, and only final messages cross the boundary. Tool calls, file edits and reasoning never reach the other side, and status messages are collected into a summary.
Why do the agents start without permission prompts?
abg claude adds --dangerously-skip-permissions and abg codex adds --yolo: without that an unattended pair could not work without a person. To restore prompts add --safe to both commands or set AGENTBRIDGE_SAFE=1. If you pass your own permission flags, the defaults switch off.
What happens when one agent runs out of its subscription limit?
abg budget shows the remaining quota of both subscription windows, five-hour and weekly. When one side hits its limit it stops at a turn boundary and hands the task to the other. The budget coordination section of the project documentation describes the automatic resume.
Discussion
No questions or reviews yet.
Related
PluginsA terminal workspace where several CLI agents like Codex, Claude and Gemini exchange messages and work on one task under human oversight
AI Builder Club Skills
AI Builder Club — Skills
A Claude Code plugin marketplace to make a repo agent-ready and run compounding agent loops
An operating layer for Claude Code and Codex: one /do entry point, project memory across sessions, safety hooks, cost telemetry and parallel agents
Multi-agent orchestration for Claude Code: teams, roles, autopilot and task routing across skills and subagents
A local real-time dashboard for Claude Code and Codex sessions: agent activity and tool calls
claude-code-sub-agents
Claude Code Subagents Collection
A collection of specialized subagents for Claude Code: frontend, backend, data, infrastructure, testing, security
Comparison with similar tools
Entries that share the main search query, and how they differ.
| Entry | Official | Repository stars | Main difference |
|---|---|---|---|
| AgentBridge: a bridge between Claude Code and Codex (this entry) | no | 416 | |
| Alook | no | 1,473 | handles and inboxes for agents of different vendors, with no standing Claude and Codex pair in one session |
| agmsg | no | 1,569 | a skill for messaging between terminal agents through a shared folder |
README in short
The bridge runs as two processes: a client inside Claude Code and a background daemon with a proxy to the Codex app server. Ports start at 4500, and a source tag on every message prevents loops by keeping it from returning to its sender.