A self-hostable middleware to manage all your MCPs through a GUI and a local proxy, supporting multiple clients and workspaces.
MetaMCP is a MCP proxy that lets you dynamically aggregate MCP servers into a unified MCP server, and apply middlewares. MetaMCP itself is a MCP server so it can be easily plugged into ANY MCP clients.
For more details, consider visiting our documentation site: https://docs.metamcp.com
English | δΈζ
Generally developers can use MetaMCP as infrastructure to host dynamically composed MCP servers through a unified endpoint, and build agents on top of it.
Quick demo video: https://youtu.be/Cf6jVd2saAs
A MCP server configuration that tells MetaMCP how to start a MCP server.
"HackerNews": {
"type": "STDIO",
"command": "uvx",
"args": ["mcp-hn"]
}
Similar to the official MCP inspector, but with saved server configs - MetaMCP automatically creates configurations so you can debug MetaMCP endpoints immediately.
Clone repo, prepare .env
, and start with docker compose:
git clone https://github.com/metatool-ai/metamcp.git
cd metamcp
cp example.env .env
docker compose up -d
If you modify APP_URL env vars, make sure you only access from the APP_URL, because MetaMCP enforces CORS policy on the URL, so no other URL is accessible.
Note that the pg volume name may collide with your other pg dockers, which is global, consider rename it in docker-compose.yml
:
volumes:
metamcp_postgres_data:
driver: local
Still recommend running postgres through docker for easy setup:
pnpm install
pnpm dev
If you have questions, feel free to leave GitHub issues or PRs.
Example mcp.json
{
"mcpServers": {
"MetaMCP": {
"url": "http://localhost:12008/metamcp/<YOUR_ENDPOINT_NAME>/sse"
}
}
}
Since MetaMCP endpoints are remote only (SSE, Streamable HTTP, OpenAPI), clients that only support stdio servers (like Claude Desktop) need a local proxy to connect.
Note: While mcp-remote
is sometimes suggested for this purpose, it's designed for OAuth-based authentication and doesn't work with MetaMCP's API key authentication. Based on testing, mcp-proxy
is the recommended solution.
Here's a working configuration for Claude Desktop using mcp-proxy
:
Using Streamable HTTP
{
"mcpServers": {
"MetaMCP": {
"command": "uvx",
"args": [
"mcp-proxy",
"--transport",
"streamablehttp",
"http://localhost:12008/metamcp/<YOUR_ENDPOINT_NAME>/mcp"
],
"env": {
"API_ACCESS_TOKEN": "<YOUR_API_KEY_HERE>"
}
}
}
}
Using SSE
{
"mcpServers": {
"ehn": {
"command": "uvx",
"args": [
"mcp-proxy",
"http://localhost:12008/metamcp/<YOUR_ENDPOINT_NAME>/sse"
],
"env": {
"API_ACCESS_TOKEN": "<YOUR_API_KEY_HERE>"
}
}
}
}
Important notes:
<YOUR_ENDPOINT_NAME>
with your actual endpoint name<YOUR_API_KEY_HERE>
with your MetaMCP API key (format: sk_mt_...
)For more details and alternative approaches, see issue #76.
?api_key=
param api key auth doesn't work for SSE. It only works for Streamable HTTP and OpenAPI.Authorization: Bearer <API_KEY>
header.uvx
or npx
, you need to customize the Dockerfile to install dependencies on your own.π οΈ Solution: Customize the Dockerfile to add dependencies or pre-install packages to reduce cold start time.
Authorization: Bearer <api-key>
headerMetaMCP supports OpenID Connect authentication for enterprise SSO integration. This allows organizations to use their existing identity providers (Auth0, Keycloak, Azure AD, etc.) for authentication.
Add the following environment variables to your .env
file:
# Required
OIDC_CLIENT_ID=your-oidc-client-id
OIDC_CLIENT_SECRET=your-oidc-client-secret
OIDC_DISCOVERY_URL=https://your-provider.com/.well-known/openid-configuration
# Optional customization
OIDC_PROVIDER_ID=oidc
OIDC_SCOPES=openid email profile
OIDC_PKCE=true
MetaMCP has been tested with popular OIDC providers:
https://your-domain.auth0.com/.well-known/openid-configuration
https://your-keycloak.com/realms/your-realm/.well-known/openid-configuration
https://login.microsoftonline.com/your-tenant-id/v2.0/.well-known/openid-configuration
https://accounts.google.com/.well-known/openid-configuration
https://your-domain.okta.com/.well-known/openid-configuration
Once configured, users will see a "Sign in with OIDC" button on the login page alongside the email/password form. The authentication flow automatically creates new users on first login.
For more detailed configuration examples and troubleshooting, see CONTRIBUTING.md.
If you want to deploy it to a online service or a VPS, a instance of at least 2GB-4GB of memory is required. And the larger size, the better performance.
Since MCP leverages SSE for long connection, if you are using reverse proxy like nginx, please refer to an example setup nginx.conf.example
Note: Prompts and resources follow similar patterns to tools.
sequenceDiagram
participant MCPClient as MCP Client (e.g., Claude Desktop)
participant MetaMCP as MetaMCP Server
participant MCPServers as Installed MCP Servers
MCPClient ->> MetaMCP: Request list tools
loop For each listed MCP Server
MetaMCP ->> MCPServers: Request list_tools
MCPServers ->> MetaMCP: Return list of tools
end
MetaMCP ->> MetaMCP: Aggregate tool lists & apply middleware
MetaMCP ->> MCPClient: Return aggregated list of tools
MCPClient ->> MetaMCP: Call tool
MetaMCP ->> MCPServers: call_tool to target MCP Server
MCPServers ->> MetaMCP: Return tool response
MetaMCP ->> MCPClient: Return tool response
Potential next steps:
See README-i18n.md
Currently en and zh locale are supported, but welcome contributions.
We welcome contributions! See details at CONTRIBUTING.md
MIT
Would appreciate if you mentioned with back links if your projects use the code.
Some code inspired by:
Not directly used the code by took ideas from
Official MCP server for Buildable AI-powered development platform. Enables AI assistants to manage tasks, track progress, get project context, and collaborate with humans on software projects.
Query the BuiltWith API to discover the technology stacks of websites. Requires a BuiltWith API key.
Create secure tunnels to expose local servers to the internet using untun.
Fetches API information from Feishu OpenAPI for seamless integration and management within an IDE.
Fetches and sets up an MCP server from a specified GitHub repository.
A secure MCP server for eBPF, designed for AI integration, kernel introspection, and automation.
An MCP server for interacting with the Tatara blockchain ecosystem. Requires configuration for the Tatara RPC endpoint and a wallet private key.
Provides linear programming (LP) and mixed-integer programming (MIP) optimization capabilities using the HiGHS solver.
ComputerVision-based πͺ sorcery of image recognition and editing tools for AI assistants.
Validate and visualize chess positions using FEN notation.