The scaling problem
The Model Context Protocol changed how AI assistants connect to external systems. But as adoption grows, a structural problem emerges: every MCP server you add registers its tool schemas with the model. Each schema consumes context tokens. At 5 servers with 20 tools each, that is roughly 15,000 tokens spent on schemas alone, on every single request. At 20 servers, the cost exceeds 60,000 tokens.
This is not sustainable. Large schema sets degrade tool selection accuracy, inflate latency, and crowd out the context window space that should be reserved for the actual work.
MetaMCP
Today, we are open-sourcing MetaMCP, a meta-server for the Model Context Protocol. MetaMCP sits between your model and any number of child MCP servers, collapsing their combined tool surface into 6 constant-cost tools at approximately 1,300 tokens, regardless of how many servers are connected.
Think of it as a power strip for MCP servers. Plug in as many as you need. Playwright, databases, GitHub, custom internal tools. Your model talks to one server. MetaMCP handles everything behind the scenes.
+--- playwright (52 tools)
|
LLM --> MetaMCP --------+--- fetch (3 tools)
(6 tools) |
+--- sqlite (6 tools)
|
+--- ... N more servers
How it works
MetaMCP exposes 6 tools to the model. Four are core operations; two provide skill awareness.
Discovery and routing. mcp_discover searches tool catalogs across all connected servers. mcp_provision accepts a natural-language description of what you need and resolves the right server, searching local catalogs first, then the npm registry for installable servers. mcp_call forwards a tool invocation to a specific server, handling connection management and crash recovery transparently.
Code execution. mcp_execute runs JavaScript in a locked-down V8 sandbox with access to all provisioned servers. This enables multi-step workflows, loops, and conditionals composed across multiple servers in a single call. The sandbox enforces 15 security protections: no eval, no require, no network access outside the server pool.
Skill awareness. mcp_skill_discover finds installed methodology files that teach agents how to use MCP servers effectively. mcp_skill_advise performs a pre-flight check, verifying that a skill's required servers are connected before invocation begins. Skills bridge what an agent can do (MCP tools) with how it should do it (methodology).
Beyond the tool surface, MetaMCP manages connection pooling with LIFO idle eviction, per-server circuit breakers with smart error classification, schema caching for fast cold starts, hot reload of configuration changes, and multi-transport support across stdio, HTTP, and SSE with OAuth.
The gallery
MetaMCP ships with a curated gallery of 122 MCP servers spanning developer tools, databases, browser automation, search, security, monitoring, and more.
metamcp add playwright sentry memory postgres
metamcp add --list
metamcp add --category databases
When you add a server that has a companion skill installed, MetaMCP detects the pairing and surfaces it. Servers and skills form a complete unit: the capability and the methodology for using it.
Auto-configuration
A single command configures MetaMCP across all major AI-powered editors:
npx @mentu/metamcp init
This detects and configures Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, Codex, and OpenCode. No manual JSON editing required.
Open source
MetaMCP is released under the Apache-2.0 license. The entire codebase, approximately 5,500 lines of TypeScript across 28 modules, is available on GitHub. Full documentation is published at metamcp.org with 22 pages covering concepts, guides, and API reference.
We believe intelligence infrastructure should be shared infrastructure. The challenges of context scaling, connection management, and tool orchestration are universal. They deserve an open solution.
Getting started
Install and run MetaMCP directly:
npx @mentu/metamcp
Or install globally:
npm install -g @mentu/metamcp
Auto-configure your editor, add servers from the gallery, and point any MCP client at MetaMCP. Full documentation at metamcp.org.