Skip to main content

Introduction

You're deep in a change with Claude Code, and you want a second opinion -- so you copy the diff, switch to a Codex or Copilot tab, paste it, wait, and copy the response back. MCP Agent Bridge removes that context switch: it exposes Claude Code, OpenAI Codex, and GitHub Copilot as Model Context Protocol (MCP) servers, so any one of them can call any other directly, from inside the same session, and get structured JSON back instead of another wall of prose to re-read.

Why?​

AI coding agents are powerful individually, but they each have blind spots. By bridging them behind MCP, you can:

  • Get a second opinion -- Send your diff to a different model for review before merging
  • Cross-validate -- Compare responses from Claude, Codex, and Copilot on the same question
  • Structured reviews -- Get typed JSON output with verdict, issues by severity, and suggestions
  • Automate workflows -- Build pipelines that route work to the best agent for each task

How It Works​

┌──────────────┐     MCP (stdio)      ┌─────────────────┐     CLI subprocess
│  MCP Client  │ ◄──────────────────► │  MCP Server     │ ──────────────────► claude -p
│  (any agent) │     or HTTP/SSE      │  (bridge layer) │                     codex exec
└──────────────┘                      └─────────────────┘                     copilot -p

Each bridge server:

  1. Registers MCP tools (review, ask, code_review, etc.)
  2. Receives tool calls from any MCP client
  3. Spawns the underlying CLI agent as a subprocess
  4. Parses and returns structured results

Architecture​

In production, each server runs as a singleton background service behind an HTTP proxy (mcp-proxy). Multiple clients share a single instance of each agent -- no cold starts, no duplicated processes.

┌─────────────┐
│ Claude Code  │──┐
└─────────────┘  │    HTTP/SSE         ┌───────────┐    stdio    ┌──────────────────┐
                 ├──────────────────► │ mcp-proxy │ ──────────► │ claude-mcp-server│
┌─────────────┐  │   :8940/mcp        └───────────┘             └──────────────────┘
│   Codex      │──┘
└─────────────┘

┌─────────────┐
│ Claude Code  │──────────────────►  :8941/mcp ──► codex-mcp-server
└─────────────┘

┌─────────────┐
│ Claude Code  │──────────────────►  :8945/mcp ──► copilot-mcp-server
└─────────────┘

Key Features​

FeatureDescription
Three agent bridgesClaude, Codex, and Copilot as MCP servers
Structured reviewsJSON output with verdict, issues, suggestions
Singleton servicesBackground daemons shared by multiple clients
Cross-platformmacOS (LaunchAgents), Windows (NSSM), Linux (systemd)
Standalone binariesZero-dependency executables compiled with Bun
Secure defaultsRead-only tools, localhost binding, input limits

Next Steps​