Agent Bridge MCP
A bridge between local CLI agents: Codex, Cursor or Claude Code delegate work to Antigravity, Grok, Kimi and seven more.
- 126 ★ on the repository
- Last commit
Broad access
Why this level
- Worker agents edit files in the project directory without a human at every step.
- The Codex CLI worker starts with `--approve-for-me`, so action approvals are off.
Install
- Works in
Manual install
uv tool install git+https://github.com/FeiZhuLulu/Agent-Bridge.gitNeeds `uv`. Installs the `agent-bridge` command; then register the server with a coordinator, for example `codex mcp add agent_bridge -- <path to agent-bridge>`.
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- Worker agents edit files in the project directory without a human at every step.
- The Codex CLI worker starts with `--approve-for-me`, so action approvals are off.
This is third-party code. Review the repository files before installing.
What it does
Agent Bridge is an MCP server through which one coordinator agent (Codex, Cursor, Kimi Code, ZCode, Grok Build, Claude Code or Devin) hands tasks to other local agents: Antigravity CLI, Grok Build, Kimi Code, DeepSeek Harness, OpenCode, Claude Code, Codex CLI, Devin CLI, ZCode and MiniMax Code. The user talks only to the coordinator; the bridge does not drive GUIs. There are ten tools. `dispatch_task` starts or resumes a turn in the project directory, `wait_task` and `check_task` track status, `get_result` returns the outcome in pages with the list of changed files, and `list_agents` lists workers and remaining quota. Coordinator modes `manual`, `auto` and `eager` set how readily it delegates.
Who it is for. Users of several CLI agents who want to drive all work from one window and hand some tasks to neighbouring agents.
Good fit when
- Research goes to Antigravity and code edits to Codex while Claude Code stays the coordinator.
- Remaining quota of Codex and Kimi Code is checked before choosing a worker.
- A long task runs in the background and the coordinator polls status through `check_task`.
- A long result is read in pages together with the list of changed files.
Not a fit when
- One agent handles every task, so the bridge adds an extra layer.
- You need to drive a graphical application: the bridge works only with CLI agents.
- Agents run on a remote machine: the server starts locally.
Example request
List agents with their remaining quota, send the research to Antigravity and the code edit to Codex.MCP
- Transport
- stdio
- Authentication
- not required
Environment variables
AGENT_BRIDGE_MODE- Coordinator mode for one client: `manual`, `auto` or `eager`.
Limitations
You need `uv`. Each worker agent is installed and signed in separately: the bridge launches their CLIs and does not replace access to them. Remaining quota is read out of the box for Codex and Kimi Code, is an experimental option for Grok Build and Claude Code, and answers `unknown` for the rest. A turn with no word from the worker for over 1800 seconds ends as `stalled`. README examples target Windows and PowerShell, so substitute paths on macOS and Linux. The Codex CLI worker runs with `--approve-for-me`.
How to disable. Remove the `agent-bridge` entry from the coordinator's MCP configuration and run `uv tool uninstall agent-bridge`. Delete the `~/.agent-bridge` directory with logs and transcripts by hand.
FAQ
Which tools does Agent Bridge have?
`list_agents`, `set_preferences`, `dispatch_task`, `wait_task`, `check_task`, `get_result`, `get_transcript`, `cancel_task`, `list_sessions` and `end_session`.
How do I add a new agent?
If the CLI speaks ACP, add an `[agents.mycustom]` block with `protocol = "acp"` and a launch command to `~/.agent-bridge/agents.toml`. Others need adapter work, described in `skills/add-worker/SKILL.md`.
How does the bridge report remaining quota?
Each `list_agents` row carries `quota` with status `ok`, `exhausted` or `unknown` and windows with `remaining_percent`. The bridge only reports; the coordinator decides routing.
How do I update the bridge?
Close coordinators holding the bridge, run `agent-bridge upgrade` and restart them.
How do I read a long result?
`get_result` returns up to 60,000 characters per call. While `has_more` is true, continue with `next_cursor`.
Discussion
No questions or reviews yet.
Related
MCP serversA local bridge linking Claude Code and Codex in one session so the agents message each other and review each other's work
Agents Can Communicate (ACC)
Agents Can Communicate
ACC passes questions, reviews and handoffs between Claude Code, Codex and other client sessions on one machine
AgentBridge by Graphene Lab
AgentBridge Graphene Lab
The standalone AgentBridge server exposes over MCP one office agent with Word, Excel, email and a task scheduler
A local dashboard for running many agents across machines: memory, code graph and agent-to-agent messaging
A tmux-based TUI that runs Claude Code, Codex and other agents in one list: status, quick prompts, worktrees and diff review
A layer over coding agents that assembles a temporary specialist team per task inside Claude Code, Codex, Cursor and Gemini
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Authentication | Repository stars | Main difference |
|---|---|---|---|---|
| Agent Bridge (this card) | no | not required | 126 | |
| AgentBridge: a bridge between Claude Code and Codex | no | 422 | agentbridge links Claude Code and Codex in one session; Agent Bridge drives ten CLI agents and reports remaining quota | |
| CCB (Claude Codex Bridge) | no | 3,609 | ccb offers a terminal workspace where agents exchange messages; Agent Bridge works as an MCP server inside the coordinator client |
README in short
The bridge keeps worker agents as processes and gives the coordinator a log and an outcome for each turn. Delegation rules live in `ORCHESTRATION.md`, and the coordinator skill is written on the first server start.