MetaMCP MCP — 使用場景, 安裝 & 即時演示

為什麼要用

核心特性

Aggregate N MCP servers → one namespaced endpoint
Multiple transports: SSE, Streamable HTTP, OpenAPI
Auth: API keys, OAuth 2.1 (MCP 2025-06-18 spec), Better Auth, OIDC SSO
Rate limits per endpoint and per user; tool overrides and renaming
Inspector UI for debugging tool calls without a real client
Docker Compose self-host in ~5 minutes

即時演示

實際使用效果

metamcp.replay ▶ 就緒

0/0

安裝

選擇你的客戶端

~/Library/Application Support/Claude/claude_desktop_config.json · Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "metamcp": {
      "command": "uvx",
      "args": [
        "mcp-proxy",
        "http://localhost:12008/metamcp/default/sse"
      ],
      "env": {
        "API_ACCESS_TOKEN": "mm_..."
      }
    }
  }
}

開啟 Claude Desktop → Settings → Developer → Edit Config。儲存後重啟應用。

~/.cursor/mcp.json · .cursor/mcp.json

{
  "mcpServers": {
    "metamcp": {
      "command": "uvx",
      "args": [
        "mcp-proxy",
        "http://localhost:12008/metamcp/default/sse"
      ],
      "env": {
        "API_ACCESS_TOKEN": "mm_..."
      }
    }
  }
}

Cursor 使用與 Claude Desktop 相同的 mcpServers 格式。專案級設定優先於全域。

VS Code → Cline → MCP Servers → Edit

{
  "mcpServers": {
    "metamcp": {
      "command": "uvx",
      "args": [
        "mcp-proxy",
        "http://localhost:12008/metamcp/default/sse"
      ],
      "env": {
        "API_ACCESS_TOKEN": "mm_..."
      }
    }
  }
}

點擊 Cline 側欄中的 MCP Servers 圖示，然後選 "Edit Configuration"。

~/.codeium/windsurf/mcp_config.json

{
  "mcpServers": {
    "metamcp": {
      "command": "uvx",
      "args": [
        "mcp-proxy",
        "http://localhost:12008/metamcp/default/sse"
      ],
      "env": {
        "API_ACCESS_TOKEN": "mm_..."
      }
    }
  }
}

格式與 Claude Desktop 相同。重啟 Windsurf 生效。

~/.continue/config.json

{
  "mcpServers": [
    {
      "name": "metamcp",
      "command": "uvx",
      "args": [
        "mcp-proxy",
        "http://localhost:12008/metamcp/default/sse"
      ]
    }
  ]
}

Continue 使用伺服器物件陣列，而非映射。

~/.config/zed/settings.json

{
  "context_servers": {
    "metamcp": {
      "command": {
        "path": "uvx",
        "args": [
          "mcp-proxy",
          "http://localhost:12008/metamcp/default/sse"
        ]
      }
    }
  }
}

加入 context_servers。Zed 儲存後熱重載。

claude mcp add metamcp -- uvx mcp-proxy http://localhost:12008/metamcp/default/sse

一行命令搞定。用 claude mcp list 驗證，claude mcp remove 移除。

使用場景

實戰用法： MetaMCP

Give a team one MCP URL instead of 15 per-laptop configs

👤 Platform teams rolling out Claude Desktop / Cursor org-wide ⏱ ~45 min intermediate

何時使用： Your team wants filesystem, github, postgres, sentry, slack, notion all available — but you don't want every engineer maintaining 6 tokens in JSON files that leak.

前置條件

A server that runs Docker — Any Linux box / VM with Docker Compose; ~1 CPU / 1GB RAM is plenty
Upstream credentials stored server-side — Put GITHUB_TOKEN, DB_URL, etc. in MetaMCP's .env — they never reach laptops

步驟

Bring up MetaMCP

git clone github.com/metatool-ai/metamcp && cp example.env .env && docker compose up -d✓ 已複製

→ Admin UI reachable at http://host:12008; default login works
Register upstream servers

In the MetaMCP admin UI, add filesystem, github, postgres, notion as namespaces 'fs', 'gh', 'db', 'notion'.✓ 已複製

→ Each appears in the endpoint's tool list with prefix
Give the team one config

Publish the Claude Desktop config snippet — one mcp-proxy entry pointing at /metamcp/team-default/sse — to internal docs.✓ 已複製

→ Every engineer paste-and-goes

結果： New hires onboard in 90 seconds; token rotation happens once, server-side.

注意事項

Claude Desktop can't do SSE natively — Use mcp-proxy (uvx mcp-proxy <sse-url>) on the client side — MetaMCP docs include this snippet
Namespaces collide or are ugly — Rename tools in MetaMCP's override UI; e.g. 'fs.read_text_file' → 'repo.read'

搭配使用： filesystem · github · postgres

Audit which MCP tools your agents actually call — and throttle the noisy ones

👤 Security / platform engineers ⏱ ~20 min intermediate

何時使用： You need to know what tools are invoked, how often, and which user. Per-server rate limits are not enough — you want aggregate visibility.

步驟

Turn on request logs

In MetaMCP settings, enable per-request logging and endpoint-level rate limits (e.g. 10 req/s per user).✓ 已複製

→ Logs appear in the admin UI
Replay a suspicious trace

Pull the last hour of tool calls for user 'marketing-bot'. Anything hitting the github write tools?✓ 已複製

→ Tabular audit; write calls highlighted
Throttle abusive tools

Set 'notion.create_page' to 1 req/min for the 'marketing-bot' endpoint. Apply now.✓ 已複製

→ Rate limit visible in config; takes effect immediately

結果： Governance surface without writing custom middleware.

Route dev/staging/prod MCP traffic through one gateway

👤 Teams with environment-specific credentials ⏱ ~30 min advanced

何時使用： You want the same Claude Desktop config to hit staging-postgres at work and prod-postgres only with an explicit flag — no editing JSON.

步驟

Create endpoints

Create 3 MetaMCP endpoints: /dev, /staging, /prod. Wire each to the right upstream credentials.✓ 已複製

→ Three endpoint URLs
Per-endpoint auth

Give prod its own API key and enable OIDC SSO for that endpoint only.✓ 已複製

→ dev/staging open with shared key, prod behind SSO
Test via inspector

Use MetaMCP's inspector UI: call a read-only query through /prod as yourself; confirm auth works.✓ 已複製

→ Successful tool call in the inspector

結果： Accident-proof environment switching.

搭配使用： postgres

組合

與其他 MCP 搭配，撬動十倍槓桿

metamcp + filesystem + github + postgres

Hide the per-tool sprawl behind one gateway URL

Configure MetaMCP with fs, gh, db namespaces; point Claude Desktop at the gateway; stop pasting tokens into JSON.✓ 已複製

metamcp + claude-desktop

Claude Desktop speaks stdio, MetaMCP speaks SSE — use mcp-proxy as the bridge

In claude_desktop_config.json, wrap MetaMCP's SSE URL with uvx mcp-proxy.✓ 已複製

metamcp + sentry

Send MetaMCP request logs/errors into Sentry for alerting

Wire MetaMCP's OpenTelemetry/logs into Sentry; alert on 5xx rate > 1%.✓ 已複製

工具

此 MCP 暴露的能力

工具	輸入參數	何時呼叫	成本
(dynamic)	Depends on which MCP servers you've registered — tools are forwarded with a namespace prefix	Any time — it's a transparent proxy with auth + rate-limit added	free (local gateway + upstream's own cost)

成本與限制

運行它的成本

API 配額: None — self-hosted
每次呼叫 Token 數: Same as upstream tools; MetaMCP adds ~1–3ms of proxy latency
費用: Free (MIT); you pay for the VM
提示: One small VPS ($5/mo) runs MetaMCP + a couple upstream servers; no need for Kubernetes until traffic > 100 req/s

安全

權限、密鑰、影響範圍

憑證儲存： MetaMCP .env holds all upstream credentials; clients never see them

資料出站： Only to the upstream MCPs you've registered. Nothing phones home.

切勿授予： Expose the admin UI to the public internet without SSO or auth proxy

The default Docker Compose binds 12008 to 0.0.0.0 — put MetaMCP behind Tailscale, Cloudflare Tunnel, or a reverse proxy with TLS before exposing it to teammates.
Tokens in .env are plain text. Mount a secrets provider (1Password CLI, Vault) for prod.

故障排查

常見錯誤與修復

Claude Desktop shows 'server failed to start'

Claude Desktop uses stdio, not SSE. Configure with uvx mcp-proxy http://host:12008/metamcp/<endpoint>/sse, not a direct URL.

驗證： Try the same URL in MetaMCP's inspector first — if it works there, the issue is stdio-to-SSE bridging

Tools missing or not namespaced

Check the endpoint config — namespaces are per-endpoint, not global. Re-sync tool discovery after adding an upstream.

驗證： In admin UI, click 'Re-sync tools' on the endpoint

401 at the gateway

API_ACCESS_TOKEN env must match the endpoint's API key; OAuth endpoints need a completed consent flow first.

驗證： curl -H 'Authorization: Bearer $TOKEN' http://host:12008/metamcp/<ep>/mcp

Upstream MCP server keeps restarting

Check the upstream container's logs via docker logs; MetaMCP will mark it unhealthy and surface errors to clients.

驗證： docker compose logs <upstream>

替代方案

MetaMCP 對比其他方案

替代方案	何時用它替代	權衡
IBM mcp-context-forge	Enterprise deployment with stronger IAM/governance needs	Heavier stack; more config surface
mcp-use/mcp-use	You want a framework to build AI agents + servers, not just a proxy	More code; more surface; more power
Direct per-client configs	You run ≤3 MCPs and don't share configs with anyone	No governance, no auth, tokens on every laptop

MetaMCP