agent-bridge
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@agent-bridgeSubmit my latest diff to the navigator for review."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
Agent Bridge MCP
A local, zero-cloud Model Context Protocol (MCP) server designed for AI-to-AI pair programming and technical peer review (Driver & Navigator).
Overview
When developing with multiple AI coding assistants (e.g., Claude Code and Google Antigravity), developers often juggle multiple terminal windows, manually copying and pasting plans, code snippets, and review critiques between models.
Agent Bridge MCP turns two local AI agents into a true pair programming duo:
Driver (Implementer): Focuses on code changes, implementation details, and refactoring.
Navigator (Architect & Auditor): Focuses on architectural integrity, edge-cases, design review, and code diff auditing.
Both agents communicate through standardized, typed MCP tools backed by an atomic, partitioned local filesystem store (.agent-bridge/) in the target project.
Related MCP server: session-coord-mcp
Key Features
Local-First & Zero-Cloud: Runs via standard
stdio. No remote servers, no databases, and no extra API keys required.Atomic & Partitioned Storage: Append-only message and proposal files (
<timestamp>-<uuid>.json) prevent data corruption and race conditions across independent operating system processes.Role-Aware Inboxes: Driver and Navigator maintain independent read cursors (
cursors/<role>.json), ensuring thatunreadqueries never clobber each other's pending items.Enforced Cross-Review: Authors cannot approve their own proposals (
review_proposalstrictly requires peer evaluation).Payload Context Guard: 500KB guard against oversized diffs, protecting the model context window.
Storage Architecture
MCP Tools Exposed
Tool Name | Purpose | Key Inputs |
| Submit a formal plan or architecture proposal for review |
|
| Issue a verdict ( |
|
| Submit code changes or a unified diff for inspection |
|
| Send a quick direct message or technical question to peer |
|
| Read unread messages, active proposals, and recent diffs |
|
Installation & Setup
Prerequisites
Node.js 18+
1. Build from source
git clone https://github.com/MPT0M/agent-bridge.git
cd agent-bridge
npm install
npm run build2. Configure Claude Code (Driver)
Add to your Claude Code MCP configuration (~/.claude/settings.json or run CLI):
claude mcp add agent-bridge node /path/to/agent-bridge/dist/index.js --role driver3. Configure Antigravity / Gemini CLI (Navigator)
Add to your Antigravity MCP settings or configuration:
{
"mcpServers": {
"agent-bridge": {
"command": "node",
"args": ["/path/to/agent-bridge/dist/index.js", "--role", "navigator"]
}
}
}Typical Workflow
Autonomous Continuous Pairing (Zero-Token Sentinel)
By default, an AI agent cannot receive unsolicited external pushes while idle. Waiting via continuous polling loops burns tokens and context quota rapidly.
agent-bridge solves this with a single-shot filesystem sentinel (agent-bridge watch). The sentinel sleeps on kernel filesystem notifications (fs.watch) using 0 tokens while waiting in silence. When peer activity is detected (a message, proposal, or diff), it coalesces rapid events (2-second window), prints a diagnostic summary, and exits with code 0. This process termination signals the agent harness to wake up automatically.
Architecture
Agent A (Active) Agent B (Sleeping)
──────────────── ──────────────────
Writes proposal or chat
│
▼
.agent-bridge/ (atomic write) ───► fs.watch kernel event
│
Coalescence window (2s)
│
agent-bridge watch exits (code 0)
│
▼
Agent B wakes up automatically!
Calls `read_bridge` (unread)
Processes peer input
Rearms watcher & goes to sleep1. In Claude Code (Driver)
When Claude finishes a turn and needs to await Navigator input, run the watcher in the background:
node /path/to/agent-bridge/dist/index.js watch --role driverIn Claude Code, running a background command emits a Background task completed notification upon process exit. Instruct Claude:
"Whenever you finish a turn waiting for Navigator feedback, run
agent-bridge watch --role driverin the background. When the background task notifies you of peer activity, callread_bridgeto inspect the inbox and rearm the watcher."
2. In Google Antigravity / Gemini CLI (Navigator)
Antigravity automatically wakes up when background tasks complete. Run the watcher as a background task:
node C:/path/to/agent-bridge/dist/index.js watch --role navigatorWhen Antigravity receives the completion message:
Calls
read_bridge({ filter: 'unread' }).Evaluates the proposal or inspects the diff.
Submits feedback via
review_proposalorchat_with_peer.Relaunches
agent-bridge watch --role navigatorin the background and ends turn.
Development & Tests
Run unit tests:
npm testBuild distribution:
npm run buildLicense
MIT © MPT0M
This server cannot be deployed
Maintenance
Related MCP Connectors
Real-time chat for AI agents. Claude Code, Cursor, Cline and Codex join channels over MCP.
Agent-native notes, tasks, dev-docs, vaults, sync & handoffs. MCP + OpenAPI dual surface.
One memory, every AI. A shared, user-owned markdown memory your AI clients read and write over MCP.
AI-native git hosting — repos, PRs, issues, CI gates, and AI code review over MCP (60 tools).
Related MCP Servers
- FlicenseNot gradedqualityAmaintenanceMCP server that bridges coding agents (Claude Code, Codex, Gemini CLI) via ACP for pair programming, enabling agents to consult each other as tools.3-
- AlicenseNot gradedqualityDmaintenanceA local-first MCP server for coordinating parallel AI coding sessions with tools like Claude Code and Codex in a single repository.2MIT
- AlicenseNot gradedqualityDmaintenanceA local MCP server that connects AI coding agents (Claude Code, Codex, Cursor, etc.) on the same machine via a shared message bus, enabling them to chat, delegate tasks, and collaborate privately without cloud or internet.54 npm18MIT
- FlicenseNot gradedqualityCmaintenanceEnables multiple AI models to collaborate under a shared goal, architecture, plan, loops, sandbox, and acceptance criteria via a local-first MCP server.-