MetaMCP

MetaMCP connects all your MCP servers through one. Your model sees 6 tools instead of hundreds.

Think of it like a power strip for MCP servers. Plug in as many as you need -- playwright, databases, GitHub, custom tools -- and your LLM talks to one server that handles everything behind the scenes.

                        ┌─── playwright (52 tools)
                        │
LLM ──► MetaMCP ────────┼─── fetch (3 tools)
        (6 tools)       │
                        ├─── sqlite (6 tools)
                        │
                        └─── ... N more servers

Cloud Run + Agent Registry Gateway

Deploy MetaMCP as a governed Cloud Run MCP gateway, register it with Google Agent Registry, and consume it from Google ADK agents with Mentu evidence for every tool call.

The production path is:

1. Turn 100 MCP tools into 6 MetaMCP gateway tools.
2. Run the gateway over Streamable HTTP on Cloud Run.
3. Mount child MCP configuration and bearer credentials from Secret Manager.
4. Register the gateway as an external MCP server in Agent Registry.
5. Resolve it from a Python ADK agent with AgentRegistry.get_mcp_toolset.
6. Export .metamcp/ledger.jsonl as a hash-linked evidence bundle.

# 1. Configure Secret Manager and the Cloud Run service account.
PROJECT_ID="your-project-id" scripts/gcp/setup-secrets.sh

# 2. Build and deploy MetaMCP to Cloud Run.
PROJECT_ID="your-project-id" scripts/gcp/deploy-cloud-run.sh

# Include INVOKER_MEMBER when an ADK runtime should call the private gateway.
PROJECT_ID="your-project-id" \
INVOKER_MEMBER="serviceAccount:[email protected]" \
scripts/gcp/deploy-cloud-run.sh

# 3. Register the deployed /mcp endpoint with Agent Registry.
PROJECT_ID="your-project-id" scripts/gcp/register-agent-registry.sh

The Agent Registry tool spec lives in gcp/agent-registry/toolspec.json, the ADK consumer sample lives in examples/gcp-adk-agent, and CI/CD examples are available for Cloud Build and Azure DevOps.

By default the Cloud Run service is deployed with --no-allow-unauthenticated. The ADK sample can send a Cloud Run identity token in X-Serverless-Authorization and the MetaMCP gateway token in Authorization.

Why MetaMCP?

Every MCP server you add registers its tool schemas with the LLM. Each schema eats context tokens. At 5 servers with 20 tools each, that's ~15,000 tokens spent on schemas alone -- every single request.

MetaMCP collapses all of that into 6 tools (~1,300 tokens). That cost stays constant whether you run 3 servers or 30. Less token overhead, better tool selection accuracy, more room for actual work.

Beyond token savings, MetaMCP handles the things you shouldn't have to think about: connection pooling, process lifecycle, error recovery, schema caching, and transport differences between local and remote servers.

Quick Start

Install and run:

npx @mentu/[email protected]        # run directly (no install)
npm install -g @mentu/metamcp    # or install globally

Auto-configure your editor (Claude Desktop, Claude Code, Cursor, VS Code, Windsurf, and more):

npx @mentu/metamcp init

Add servers from the built-in gallery (122 curated servers):

metamcp add playwright sentry memory postgres    # one-click, writes .mcp.json
metamcp add --list                               # browse all available servers
metamcp add --category search                    # filter by category

Or create a .mcp.json manually:

{
  "mcpServers": {
    "playwright": {
      "command": "npx",
      "args": ["-y", "@playwright/mcp@latest"]
    },
    "sqlite": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sqlite", "/path/to/db"]
    }
  }
}

metamcp --config .mcp.json

That's it. MetaMCP speaks MCP over stdio for local clients and Streamable HTTP for remote clients such as Cloud Run.

Run remote HTTP mode locally:

npm run build
METAMCP_TRANSPORT=http METAMCP_CONFIG=.mcp.json PORT=8080 npm start
curl http://localhost:8080/healthz

Note: MetaMCP optionally uses better-sqlite3 for semantic search. This requires a C++ compiler. If compilation fails, MetaMCP still works with keyword-only search. On macOS: xcode-select --install. On Linux: apt install build-essential.

The Tools

MetaMCP gives the LLM 6 tools: 4 core tools for server management, and 2 advisory tools for skill awareness.

Core Tools

mcp_discover -- Find tools

Search tool catalogs across all connected servers. Without a query, returns server status and tool counts.

{ "query": "screenshot" }

mcp_provision -- Get what you need

Describe a capability and MetaMCP resolves the right server. It searches local catalogs first, then the npm registry for installable servers.

{ "intent": "I need to crawl a website and extract links" }

mcp_call -- Use a tool

Forward a tool call to a specific server. MetaMCP handles connection management and retries on crash.

{ "server": "playwright", "tool": "browser_navigate", "args": { "url": "https://example.com" } }

mcp_execute -- Write code

Run JavaScript in a V8 sandbox with access to all provisioned servers. Compose multi-step workflows, loops, and conditionals in a single call.

{ "code": "const result = await servers.sqlite.call('query', { sql: 'SELECT count(*) FROM users' }); return result;" }

Skill-Aware Tools

Skills are methodology files (SKILL.md) that teach agents how to use MCP servers effectively. MetaMCP can discover skills and check whether their required MCP servers are available.

mcp_skill_discover -- Find skills

Search installed skills with MCP readiness status. Returns matching skills, their required servers, and whether those servers are connected.

{ "query": "browser automation" }

mcp_skill_advise -- Pre-flight check

Check whether a specific skill's dependencies are satisfied before using it.

{ "skill": "playwright" }

Skills live in ~/.claude/skills/ (personal) or .claude/skills/ (project). MetaMCP scans both locations and matches skills to their companion MCP servers via the requires-mcp frontmatter field. See Skills in the docs.

Server Gallery

MetaMCP ships with a curated gallery of 122 MCP servers across developer tools, databases, browser automation, search, security, monitoring, and more.

metamcp add --list                    # browse all servers
metamcp add playwright sentry neon    # add multiple at once
metamcp add --category databases      # filter by category

When you add a server that has a companion skill installed, MetaMCP tells you:

Added 2 server(s): playwright, sentry

Companion skills detected:
  playwright → skill: playwright
  sentry → skill: sentry
These skills teach agents how to use these servers effectively.

See the full gallery at metamcp.org/guides/server-gallery.

Evidence Export

MetaMCP records tool activity in .metamcp/ledger.jsonl. You can export that append-only ledger into a hash-linked evidence bundle:

metamcp export-evidence --ledger .metamcp/ledger.jsonl --out .metamcp/evidence-bundle.json
metamcp export-evidence --out .metamcp/evidence-bundle.json --verify

The bundle schema is io.mentu.metamcp.evidence-bundle.v1. Each entry includes the original ledger event, its previous hash, and its current SHA-256 hash so later systems can verify continuity.

Configuration

MetaMCP reads .mcp.json -- the same format used by Claude Desktop and Claude Code.

Local server:

{
  "mcpServers": {
    "my-server": {
      "command": "/usr/local/bin/my-mcp-server",
      "args": ["--port", "8080"],
      "env": { "API_KEY": "..." }
    }
  }
}

Remote server (SSE):

{
  "mcpServers": {
    "remote-tools": {
      "url": "https://mcp.example.com/sse",
      "transportType": "sse",
      "headers": { "Authorization": "Bearer your-token" }
    }
  }
}

Remote server (HTTP) with OAuth:

{
  "mcpServers": {
    "cloud-server": {
      "url": "https://mcp.example.com/api",
      "oauth": true
    }
  }
}

Server lifecycle:

{
  "mcpServers": {
    "database": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-sqlite", "/path/to/db"],
      "lifecycle": { "mode": "keep-alive", "idleTimeoutMs": 600000 }
    },
    "one-shot": {
      "command": "/usr/local/bin/converter",
      "lifecycle": "ephemeral"
    }
  }
}

Three transport types: stdio (local, default), http (Streamable HTTP), and sse (Server-Sent Events). OAuth triggers a browser flow on first connect, with tokens saved to ~/.metamcp/oauth/.

Lifecycle controls idle behavior: keep-alive servers persist, ephemeral servers tear down immediately after use, and servers without a declaration follow the default pool timeout.

Secret resolution from the vault

Hard-coding API_KEY strings in .mcp.json is the easiest way to leak credentials into git. MetaMCP supports ${KEY} references in any env or headers value and resolves them at config load time:

{
  "mcpServers": {
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": { "GITHUB_TOKEN": "${GITHUB_TOKEN}" }
    },
    "remote-tools": {
      "url": "https://mcp.example.com/sse",
      "transportType": "sse",
      "headers": { "Authorization": "Bearer ${REMOTE_TOOLS_TOKEN}" }
    }
  }
}

Resolution order for each ${KEY}:

1. mentu vault - if mentu-vault is installed at ~/.local/bin/mentu-vault, MetaMCP looks up the key in the macOS Keychain (or the age-encrypted file fallback). Workspace-scoped lookup is tried first when MENTU_WORKSPACE is set, then global.
2. process.env - standard environment variable.
3. Literal - if neither resolves, MetaMCP logs a warning and leaves the ${KEY} reference in place so misconfiguration is visible instead of silent.

Vault lookups are cached for the process lifetime, so resolution happens once at startup with no per-connection overhead. Inline references like "Bearer ${TOKEN}" and standalone "${TOKEN}" are both supported.

If you do not use mentu vault, MetaMCP falls back to process.env automatically - no extra config needed. Just export GITHUB_TOKEN=... and the same .mcp.json works.

Secret scrubbing on the way out

MetaMCP also runs an output scrubber on every tool response before it returns to the LLM. JWTs, OpenAI/GitHub/Slack/AWS tokens, and JSON-shaped credential keys (password, secret, api_key, access_token, private_key, authorization, etc.) are replaced with [REDACTED:LABEL]. This is best-effort defense in depth - secrets that match well-known patterns get caught even if a misconfigured downstream server echoes them in an error message.

What MetaMCP handles for you

• Connection pool -- bounded pool with LIFO idle eviction. Servers start lazily on first use.
• Circuit breaker -- per-server failure tracking. Errors are classified: auth failures (401/403) never trip the breaker, only transient errors count.
• Schema caching -- tool schemas persist to disk for fast cold starts. Stale caches refresh transparently.
• Config import -- --import discovers servers from Cursor, Claude Desktop, Claude Code, VS Code, Windsurf, Codex, and OpenCode.
• Hot reload -- MetaMCP watches .mcp.json for changes. Add servers with metamcp add and they become available within 2 seconds, no restart needed.
• V8 sandbox -- mcp_execute runs in a locked-down context. No eval, no require, no network access.
• Multi-transport -- stdio, HTTP, and SSE with OAuth. The model doesn't know the difference.
• Skill awareness -- discovers companion skills for MCP servers and checks readiness before invocation.

CLI

Documentation

Full docs at metamcp.org.

Contributing

See CONTRIBUTING.md for development setup and guidelines.

Links

• Documentation
• npm
• GitHub

License

Apache-2.0