mcp-tap

The last MCP server you install by hand.

mcp-tap lives inside your AI assistant. Ask it to find, install, and configure any MCP server — by talking to it. No more editing JSON files. No more Googling environment variables. No more "why won't this connect?"

"Find me an MCP for PostgreSQL."

That's it. mcp-tap searches the registry, installs the package, generates the config, validates the connection — all through conversation.

Before mcp-tap

1. Google "MCP server for postgres"

2. Find 4 competing packages, compare stars and last commit dates

3. Pick one, read the README

4. Figure out the right command, args, and env values

5. Manually edit claude_desktop_config.json (or mcp.json, or mcp_config.json...)

6. Realize you need a POSTGRES_CONNECTION_STRING environment variable

7. Find your connection string, add it to the config, restart the client

8. Get "connection refused", debug for 20 minutes

9. Finally works. Repeat for every server. Repeat for every client.

Related MCP server: mcp-server-mcpindex

After mcp-tap

You: "Set up MCP servers for my project."

mcp-tap: I scanned your project and found:
  - PostgreSQL (from docker-compose.yml)
  - Slack (SLACK_BOT_TOKEN in your .env)
  - GitHub (detected .github/ directory)

  I recommend 3 servers. Want me to install them?

You: "Yes, all of them."

mcp-tap: Done. All connections verified. 35 new tools available.

Install

You install mcp-tap once. It installs everything else.

Claude Desktop

Add to your claude_desktop_config.json:

With uvx (recommended):

{
  "mcpServers": {
    "mcp-tap": {
      "command": "uvx",
      "args": ["mcp-tap"]
    }
  }
}

With npx:

{
  "mcpServers": {
    "mcp-tap": {
      "command": "npx",
      "args": ["-y", "mcp-tap"]
    }
  }
}

Claude Code

# With uvx (recommended)
claude mcp add mcp-tap -- uvx mcp-tap

# With npx
claude mcp add mcp-tap -- npx -y mcp-tap

Cursor

Add to .cursor/mcp.json:

{
  "mcpServers": {
    "mcp-tap": {
      "command": "uvx",
      "args": ["mcp-tap"]
    }
  }
}

Or use npx — replace "command": "uvx", "args": ["mcp-tap"] with "command": "npx", "args": ["-y", "mcp-tap"].

Windsurf

Add to ~/.codeium/windsurf/mcp_config.json:

{
  "mcpServers": {
    "mcp-tap": {
      "command": "uvx",
      "args": ["mcp-tap"]
    }
  }
}

Or use npx — replace "command": "uvx", "args": ["mcp-tap"] with "command": "npx", "args": ["-y", "mcp-tap"].

What can it do?

Tools

Plus automatic lockfile management on every configure/remove.

Features

• Project-aware: Scans your codebase — including CI/CD configs (GitHub Actions, GitLab CI) — to recommend servers based on your actual stack

• Security gate: Blocks suspicious install commands, archived repos, and known-risky patterns before installing

• Lockfile: mcp-tap.lock tracks exact versions and hashes of all your MCP servers. Reproducible setups across machines

• Stacks: Shareable server profiles — install a complete Data Science, Web Dev, or DevOps stack in one command

• Multi-client: Configure Claude Desktop, Claude Code, Cursor, and Windsurf — all at once or individually

• Auto-healing: Failed connections are automatically diagnosed and fixed when possible

• Tool conflict detection: Warns when two servers expose overlapping tools that could confuse the LLM

• Connection validation: Every install is verified with a real MCP connection test

• Secrets masked: list_installed never exposes environment variable values

• Recommendation quality gate: Offline benchmark (precision@k, acceptance_rate) keeps recommendation quality stable in CI

• Production feedback loop (opt-in): Privacy-safe telemetry (recommendations_shown, accepted/rejected/ignored) with version-segmented quality trends

• Semantic rerank: Broad queries (for example error monitoring) are reranked by intent match, not only popularity

• Offline-friendly search fallback: Uses recent cached registry results when live providers fail, with explicit staleness metadata

• Configure preflight: configure_server(dry_run=true) validates install/connectivity without writing client config

Requirements

• Python 3.11+ with uv (recommended), or

• Node.js 18+ (the npm package is a thin wrapper that calls the Python package via uvx/pipx)

• Officially tested in CI on Python 3.11, 3.12, 3.13, and 3.14

Quality Gate

Run the recommendation benchmark locally:

uv run python -m mcp_tap.benchmark.recommendation

Dataset: src/mcp_tap/benchmark/recommendation_dataset_v1.json.

Production Feedback Loop (Opt-In)

Enable telemetry explicitly before collecting production recommendation feedback:

export MCP_TAP_TELEMETRY_OPT_IN=true
export MCP_TAP_TELEMETRY_FILE=.mcp-tap/recommendation_feedback.jsonl

Generate a quality report from collected events:

uv run python -m mcp_tap.benchmark.production_feedback --events .mcp-tap/recommendation_feedback.jsonl --top-k 3

The telemetry payload is privacy-safe by default:

• project path is stored as hash fingerprint (no raw path)

• no source code, secrets, or env var values are recorded

• release trends include drift warnings/failures between versions

License

MIT