Building an MCP Bridge for OpenClaw

The Problem

OpenClaw agents are powerful, but they have a blind spot: they can't talk to MCP servers. The Model Context Protocol is an open standard that lets AI tools access structured data — documents, code, databases, APIs — through a unified interface.

Without MCP support, agents resort to brute-force file reads. Need one section from an 82KB spec? You're dumping the whole thing into the context window. That's ~20,000 tokens burned to find a paragraph.

OpenClaw's native MCP client support is on the roadmap but hasn't shipped yet. We didn't want to wait.

The Solution

We built openclaw-mcp-bridge — a Python CLI that bridges the gap. It spawns MCP servers as child processes and routes tool calls via JSON-RPC 2.0 over stdio. No daemons, no ports, no config servers.

Agents call it through the exec tool, just like any other CLI command.

Agent

OpenClaw

→

Exec

Shell

→

Bridge

mcp-bridge

→

Stdio

MCP Server

→

Data

Result

The bridge is generic — it works with any MCP server that supports stdio transport. Configure your servers in servers.json, and every agent on the machine can use them.

The Numbers

We tested with jDocMunch, an MCP server that indexes documentation by heading hierarchy. We pointed it at our 82KB project spec (1,936 lines, 552 sections across 26 files).

Without bridge

~20,000

Agent reads the entire spec to find the database schema section. Most tokens wasted.

With mcp-bridge

~84

Search by query, retrieve the exact section. Byte-precise extraction.

99.6%

Token reduction on document lookups

The indexing took 394ms. The sections persist locally, so subsequent lookups are instant. For a fleet of 6 agents that need to reference the same spec, this translates to real cost savings on every cold start.

Quick Start

Install

# Install the bridge + a doc server
pip install mcp jdocmunch-mcp
git clone https://github.com/cooeyau/openclaw-mcp-bridge.git
ln -s $(pwd)/openclaw-mcp-bridge/src/mcp_bridge.py /usr/local/bin/mcp-bridge

Index Your Docs

mcp-bridge call jdocmunch index_local \
  --args '{"path": "/your/project"}'
# → 552 sections indexed in 394ms

Search & Retrieve

# Search by natural language
mcp-bridge call jdocmunch search_sections \
  --args '{"repo":"local/project","query":"auth flow"}'

# Get just that section
mcp-bridge call jdocmunch get_section \
  --args '{"repo":"local/project","section_id":"..."}'

Compatible Servers

Any MCP server that speaks stdio plugs straight in. Here are a few we've tested:

📄 jDocMunch

Navigate docs by heading hierarchy. Retrieve exact sections instead of whole files.

pip install jdocmunch-mcp

🔧 jCodeMunch

Search code by symbol — functions, classes, methods via tree-sitter AST parsing.

pip install jcodemunch-mcp

📊 Any MCP Server

Databases, APIs, knowledge bases — if it speaks MCP over stdio, it works.

servers.json

🔌 Build Your Own

The bridge handles the protocol. You handle the logic.

modelcontextprotocol.io

What's Next

We're watching OpenClaw #4834 for native MCP client support. When it lands, agents won't need the bridge at all — but until then, this gives the community a working solution today.

The bridge is open source under MIT. Fork it, extend it, break it.

Credits: Inspired by openclaw-mcp by Tomáš Grasl — an MCP server that exposes OpenClaw to Claude.ai. We built the reverse: an MCP client that gives OpenClaw agents access to MCP servers. Uses jDocMunch and jCodeMunch by J. Gravelle for structured retrieval.