mcp-meta-hub

A lightweight MCP proxy that aggregates multiple MCP servers into two meta-tools.

Stop giving your AI agent a shell. Give it skills.

The Problem

When you connect multiple MCP servers to an AI agent, every tool from every server lands in the LLM's context window. With 10 servers × 20 tools each, that's 200 tool descriptions competing for the model's attention. The result: bloated prompts, confused tool selection, higher costs, and hard limits from clients like Cursor (80 tools max).

The Solution

mcp-meta-hub sits between your AI agent and your MCP servers. Instead of exposing all 200 tools, it exposes exactly two:

Skills (SKILL.md files) tell the agent exactly which tools to call — so in daily use, the agent goes straight to call_tool. list_tools is primarily used when creating new skills, to explore what's available.

AI Agent (Claude, Cursor, Copilot…)
    │
    │  sees 2 tools
    ▼
  mcp-meta-hub
    │
    │  connects to N servers
    ▼
  ┌─────────┬──────────┬───────────┐
  │ weather │ database │ github    │  ← MCP servers
  │ 3 tools │ 8 tools  │ 25 tools  │
  └─────────┴──────────┴───────────┘

Quick Start

Install

npm install -g mcp-meta-hub

Configure

Create mcp-hub.json:

{
  "servers": {
    "weather": {
      "command": "node",
      "args": ["./skills/weather/index.js"]
    },
    "github": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-github"],
      "env": {
        "GITHUB_TOKEN": "ghp_..."
      }
    }
  }
}

Run

mcp-meta-hub ./mcp-hub.json

Connect to your AI client

Add mcp-meta-hub to your client's MCP configuration. For example, in Claude Desktop:

{
  "mcpServers": {
    "hub": {
      "command": "mcp-meta-hub",
      "args": ["./mcp-hub.json"]
    }
  }
}

That's it. Your agent now sees 2 tools instead of 200.

How It Works

The Two Loops

Skill creation (one-time): use list_tools to explore servers, then write a SKILL.md documenting the key tools and workflows.

Daily usage (99% of the time): the agent reads a SKILL.md and calls call_tool directly — no discovery needed.

┌─────────────────────────────────────────┐
│  Skill Creator (one-time)               │
│  list_tools("github") → write SKILL.md  │
└──────────────────┬──────────────────────┘
                   │ produces
                   ▼
              SKILL.md
                   │ read by
                   ▼
┌─────────────────────────────────────────┐
│  Agent (daily usage)                    │
│  read SKILL.md → call_tool() directly   │
└─────────────────────────────────────────┘

Tool Naming Convention

Tools are namespaced as {server}__{tool}:

• weather__get_forecast
• weather__list_cities
• github__create_issue
• github__list_repos
• database__query

This prevents name collisions and makes prefix-based discovery natural.

Configuration Reference

Stdio Server (local process)

{
  "servers": {
    "my-skill": {
      "command": "node",
      "args": ["./path/to/server.js"],
      "env": {
        "API_KEY": "secret"
      }
    }
  }
}

SSE Server (remote)

{
  "servers": {
    "remote-api": {
      "url": "http://localhost:3001/sse",
      "transport": "sse"
    }
  }
}

Streamable HTTP Server (remote)

{
  "servers": {
    "remote-api": {
      "url": "http://localhost:3001/mcp",
      "transport": "streamable-http"
    }
  }
}

Skills

A skill is a SKILL.md file that tells the AI agent what tools are available and when to use them.

Skills are views over your tools, not 1:1 mappings to servers. You can slice them however you want:

• 1 skill per server — document all tools from a GitHub MCP → github skill
• 1 skill per use case — cherry-pick tools across servers → incident-response skill uses GitHub + Sentry + PagerDuty
• Partial coverage — a server has 25 tools but you only need 3? Document just those 3. The rest stay discoverable via list_tools but the agent knows exactly which ones matter for this skill.
• Multiple skills, same server — a GitHub MCP can feed a code-review skill AND a release-management skill, each with different workflows and different subsets of tools.

Think of it like SQL views: same underlying tables, different perspectives for different purposes.

The Skill Creator Pattern

The skill-creator is a special meta-skill: it's the only skill that uses list_tools. All other skills go straight to call_tool.

The workflow:

1. Agent receives "create a skill for X"
2. Skill-creator SKILL.md is activated
3. Agent calls list_tools to discover available tools and schemas
4. Agent writes a new SKILL.md with the right tools, params, and workflows
5. From now on, the new skill never calls list_tools — it goes direct

This is why mcp-meta-hub ships with a skill-creator example — it's the bootstrap skill that creates all others.

There are two types of skills:

Type 1: Documentation Skills (99% of cases)

You use existing MCP servers (from npm, the community, your team) and write a SKILL.md to give the agent context. No code required. Just markdown.

skills/devops/
└── SKILL.md       ← That's it. Zero code.

Example: Connect 3 existing MCP servers (GitHub, Sentry, PagerDuty) and let the skill-creator build the skill that matches your exact workflow:

---
name: incident-response
description: Handle production incidents. Use when the user reports a bug, outage, or needs to investigate an error in production.
---

# Incident Response

When an incident is reported:

1. Check recent errors → `call_tool("sentry__list_issues", { "status": "unresolved" })`
2. Find related PRs → `call_tool("github__search_commits", { "query": "fix OR hotfix", "since": "24h" })`
3. Page on-call if P0 → `call_tool("pagerduty__get_on_call", {})`

## Key Tools

| Tool | Description |
|---|---|
| `sentry__list_issues` | List errors. Params: `status` (unresolved/resolved), `sort` (date/freq). |
| `sentry__get_issue` | Get error details + stack trace. Params: `issueId`. |
| `github__search_commits` | Search recent commits. Params: `query`, `since`. |
| `github__create_issue` | Create tracking issue. Params: `repo`, `title`, `labels`. |
| `pagerduty__get_on_call` | Get current on-call engineer. |

## Escalation Rules

- P0 (site down): Page on-call immediately
- P1 (feature broken): Create GitHub issue, notify #incidents
- P2 (degraded): Create GitHub issue, triage next standup

The skill doesn't implement anything — it orchestrates existing tools with domain knowledge. The agent reads this and knows exactly which tools to call, with which parameters, in which order.

This is the power: your expertise becomes a reusable skill file.

Type 2: Custom MCP Servers (when no existing server fits)

When you need to wrap a CLI, an internal API, or build something custom, create an MCP server and register it in mcp-hub.json:

skills/weather/
├── SKILL.md          ← Describes the skill for the AI agent
├── src/index.ts      ← MCP server (use @modelcontextprotocol/sdk)
└── package.json

{
  "servers": {
    "weather": {
      "command": "node",
      "args": ["./skills/weather/dist/index.js"]
    }
  }
}

See examples/weather/ for a complete working example with SKILL.md + MCP server code.

Multi-Instance Skills

The same MCP server can run multiple times with different configs — useful for multi-account scenarios:

{
  "servers": {
    "accounting-company-a": {
      "command": "node",
      "args": ["./skills/accounting/dist/index.js"],
      "env": { "API_KEY": "key-for-company-a" }
    },
    "accounting-company-b": {
      "command": "node",
      "args": ["./skills/accounting/dist/index.js"],
      "env": { "API_KEY": "key-for-company-b" }
    }
  }
}

Tools become accounting-company-a__get_balance and accounting-company-b__get_balance — same skill, isolated contexts.

Advanced Configuration

Environment Variable References

Use $VAR in server env values to reference variables from the parent process environment. Useful for multi-instance setups where the same server needs different credentials:

{
  "servers": {
    "accounting-a": {
      "command": "node",
      "args": ["./accounting-server.js"],
      "env": { "DB_PASSWORD": "$ACCT_A_PASSWORD" }
    },
    "accounting-b": {
      "command": "node",
      "args": ["./accounting-server.js"],
      "env": { "DB_PASSWORD": "$ACCT_B_PASSWORD" }
    }
  }
}

The hub resolves $ACCT_A_PASSWORD from process.env at startup. If a referenced variable is not found, a warning is logged and the literal $VAR string is kept. Servers also inherit all parent environment variables automatically (unlike the MCP SDK default which only inherits HOME, PATH, etc.).

Tool Prefix

By default, tools are namespaced as {server}__{tool} to prevent collisions. You can override this per server:

{
  "servers": {
    "weather": {
      "command": "node",
      "args": ["./weather.js"],
      "prefix": true
    },
    "brave": {
      "command": "npx",
      "args": ["brave-search"],
      "prefix": false
    },
    "internal": {
      "command": "node",
      "args": ["./internal.js"],
      "prefix": "myapp__"
    }
  }
}

Why Not Just Use Bash?

The current trend is giving AI agents shell access. It's powerful but dangerous:

mcp-meta-hub enforces the principle of least privilege: the agent can only do what your skills explicitly allow.

Comparison with Existing Solutions

The key difference: other proxies aggregate tools and dump them all into the LLM context. mcp-meta-hub makes them discoverable on demand. This is the difference between loading every page of a website at once vs. having a search bar.

Development

git clone https://github.com/verdier/mcp-meta-hub.git
cd mcp-meta-hub
npm install
npm run build

# Build the example skill
cd examples/weather && npm install && npm run build && cd ../..

# Run tests
npm test

License

MIT — see LICENSE.