mcp-toolmux
Proxy-style MCP tool multiplexer that aggregates multiple downstream stdio MCP servers into one, offering meta-tools for status, search, call, parallel, batch, and pipeline operations with concurrency control and caching.
README
mcp-toolmux
mcp-toolmux is an independent MCP proxy-style tool multiplexer that exposes a small set of meta-tools and routes calls to downstream MCP servers with concurrency control, caching, and structured partial-failure handling.
It is inspired by callmux, but it is not a full clone. The current release focuses on a practical proxy shape: downstream registry, exposure modes, meta-tools, direct qualified calls, parallel calls, batch calls, dependent pipelines, tool search, schema compression, status, and lightweight caching.
中文说明:这是一个独立的 MCP proxy 项目,用来把多个下游 stdio MCP server 汇聚成一个上游可调用的 MCP server。本项目不是简单 demo,而是一个独立的 MCP 工具调用多路复用项目,目标是逐步完善工具路由、并发调度、批处理、流水线、缓存和协议层优化能力。
Architecture
flowchart LR
Client["Codex / Claude MCP client"] --> Proxy["mcp-toolmux stdio MCP server"]
Proxy --> Meta["Meta-tools\nstatus / search / call / parallel / batch / pipeline"]
Meta --> Registry["Downstream registry\nqualified + unqualified names"]
Registry --> Search["Tool search\ncompressed summaries"]
Meta --> Cache["Read-tool cache\nTTL + max entries"]
Registry --> Fake["fake stdio MCP server"]
Registry --> Other["other trusted stdio MCP servers"]
Features
- stdio MCP proxy server
- downstream MCP server registry from config
- tool exposure modes:
standardandmeta-only - meta-tools:
toolmux_status,toolmux_search_tools,toolmux_call,toolmux_parallel,toolmux_batch,toolmux_pipeline - direct qualified tool calls in standard mode, using names like
server__tool - unqualified tool calls when the tool name is not ambiguous
- global concurrency limit for parallel calls
toolmux_batchfor multiple argument sets against the same downstream tooltoolmux_pipelinefor serial dependent tool chains- input mapping with
$previous,$previous.result, and$steps.N.result toolmux_search_toolsfor downstream tool discovery- schema compression for compact tool summaries
- partial failure handling with
failedIndexes - lightweight TTL cache for read-style tool names
toolmux_parallel is for independent tool calls.
toolmux_batch is for multiple argument sets against the same tool.
toolmux_pipeline is for dependent tool chains where later steps consume earlier outputs.
Tool Exposure Modes
standard exposes meta-tools and qualified downstream tools. This mode is useful when you want transparent tool lists and direct calls such as fake__echo.
meta-only exposes only toolmux_* meta-tools. Downstream tools are discovered through toolmux_search_tools and invoked through toolmux_call, toolmux_parallel, toolmux_batch, or toolmux_pipeline.
Tool Search and Compression
toolmux_search_tools searches downstream tools by tool name, qualified name, description, and server name. It returns compressed summaries instead of full tool definitions by default.
Compressed tool summaries include:
nameserverqualifiedName- truncated
description requiredinputProperties- optional
inputSchema
Not Implemented Yet
- HTTP/SSE listener
- auth, dashboard, and metrics
- advanced policy controls
- persistent cache
The project is still evolving and does not yet include HTTP/SSE listener, auth, dashboard, advanced policy controls, or persistent cache.
Install
npm install
Build
npm run build
CLI Usage
npm install
npm run build
node dist/cli.js --help
node dist/cli.js init --output mcp-toolmux.config.json
node dist/cli.js doctor --config examples/mcp-toolmux.config.json
node dist/cli.js tools --config examples/mcp-toolmux.config.json
node dist/cli.js serve --config examples/mcp-toolmux.config.json
serve starts the stdio MCP server mode. Logs are written to stderr so MCP protocol messages on stdout stay clean.
doctor checks config shape, defaults, command availability, and common mistakes before you start the proxy.
tools briefly connects to downstream MCP servers and prints compressed downstream tool summaries as JSON.
init creates a starter standard-mode config and refuses to overwrite an existing file.
Run the Fake Server
The fake server is meant to be launched by the proxy after build:
npm run build
node dist/examples/fake-mcp-server.js
Run the Proxy
Use the example config:
npm run build
node dist/src/index.js --config examples/mcp-toolmux.config.json
The proxy always exposes six MCP meta-tools:
toolmux_statustoolmux_search_toolstoolmux_calltoolmux_paralleltoolmux_batchtoolmux_pipeline
In standard mode it also exposes qualified downstream tools such as fake__echo.
Example Config
{
"mode": "standard",
"schemaCompression": {
"enabled": true,
"descriptionMaxLength": 180,
"includeInputSchema": false,
"includeRequired": true
},
"servers": {
"fake": {
"command": "node",
"args": ["dist/examples/fake-mcp-server.js"]
}
},
"maxConcurrency": 5,
"cache": {
"enabled": true,
"ttlSeconds": 300,
"maxEntries": 256
}
}
By default, mcp-toolmux reads mcp-toolmux.config.json from the current directory. Pass --config path to use another file.
Tests
npm test
The handler tests use a mock upstream manager to keep the test suite fast. The fake MCP server remains available for manual stdio testing.
Security Note
Only configure trusted MCP server commands. mcp-toolmux never accepts arbitrary commands from tool input, and downstream server commands must come from the config file. Treat downstream MCP servers as executable code with your local privileges.
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.