mcp-lookup

Lookup MCP servers across the official MCP Registry, Smithery, and PulseMCP from one CLI and Python library. Stdlib-only, no runtime dependencies.

Why

Tools that need to enrich raw MCP server detections — e.g. a security dashboard that catalogs which MCP servers developers are running, an IDE assistant that needs canonical install metadata, or an SBOM pipeline — typically resort to HTML-scraping third-party catalogs because most registries don't expose a clean JSON API.

mcp-lookup wraps the three main MCP catalogs (registry.modelcontextprotocol.io, api.smithery.ai, and api.pulsemcp.com) behind a tiny, stable interface. It gives you:

• Canonical server metadata by name or repo URL: description, publisher namespace, version, packages (npm/pypi/docker), and remote transports (stdio/sse/streamable-http).
• Three-source aggregation: search fans out to all three sources in parallel and returns merged results, each labeled by _source. get prefers the registry (canonical), then falls back to Smithery, then PulseMCP.
• Parallel search for low latency: search latency is bounded by the slowest source (~1.2s) instead of the sum of all three (~5s). One slow or unavailable source never blocks the others.
• Authoritative tool lists: Smithery entries include the server's tools[] with input schemas, which the registry does not expose.
• Community coverage: PulseMCP adds GitHub star counts, package download counts, and long-tail community servers not (yet) in the official registry.
• URL → name resolution: hand it a https://github.com/owner/repo (or a vendor domain) and every source tries to find the matching entry.
• Verified publisher signal for free: registry namespace prefixes (io.github.<user>/…, <domain>/…) are cryptographically verified; Smithery entries carry a verified flag.
• Icon derivation: returns a usable icon URL for every server — Smithery's own iconUrl when available, GitHub owner avatars for repo-backed servers, DuckDuckGo favicons for vendor-hosted remotes.
• Subprocess-friendly: designed to be called from a backend service to replace fragile HTML scrapers. Outputs stable JSON with --json.

Typical callers:

• Dev-tool discovery agents that detect local MCP configs and want enrichment
• Security / DLP platforms classifying MCP servers by publisher and risk
• Registry-aware IDE extensions
• CI jobs validating MCP server manifests

Install

pip install mcp-lookup

Or from source:

git clone <repo-url>
cd mcp-lookup
pip install -e .

CLI

# Search across all three sources (in parallel)
mcp-lookup search github --limit 5
mcp-lookup search https://github.com/owner/repo
mcp-lookup search github --json                     # machine-readable

# Get a single record (registry → Smithery → PulseMCP fallback)
mcp-lookup get app.linear/linear --pretty
mcp-lookup get https://github.com/owner/repo --pretty
mcp-lookup get ai.smithery/<name> --version 1.2.3

# Restrict to one source
mcp-lookup --source registry search linear
mcp-lookup --source smithery get linear --pretty    # includes tools[] with descriptions
mcp-lookup --source pulsemcp search linear

# Icon derivation
mcp-lookup icon app.linear/linear
mcp-lookup get app.linear/linear --with-icon        # injects icon_url into JSON output

Exit codes: 0 success, 1 no results, 2 registry/network error.

Library

from mcp_lookup import (
    Aggregator,
    RegistryClient,
    SmitheryClient,
    PulseMCPClient,
    derive_icon_url,
)

# Combined (default) — parallel across all three sources
agg = Aggregator()
results = agg.search("linear", limit=5)
entry = agg.get("app.linear/linear")                # registry first, Smithery/PulseMCP fallback

# Single source
registry_only = agg.get("app.linear/linear", source="registry")
smithery_only = agg.get("linear", source="smithery")    # includes tools[] list
pulsemcp_only = agg.search("linear", source="pulsemcp")

# Direct clients
RegistryClient().get("app.linear/linear")
SmitheryClient().get("linear")
PulseMCPClient().search("linear", limit=3)

# Icon helper
icon_url = derive_icon_url(entry)

Contributing

Contributions are very welcome — this is an open project and we'd love help from the MCP community.

Good first areas to contribute:

• Additional sources (mcp.so, Glama.ai, GitHub topic fallback) behind the existing provider interface
• Live tools/list probing so registry-only entries can return authoritative tool lists
• Local caching (SQLite or file-based) with TTL + --refresh
• Cross-source deduping by normalized repository URL
• Typed response schemas (Pydantic / JSON Schema) for enterprise consumers
• Tests and CI coverage

To contribute:

1. Fork the repo and create a feature branch.
2. Keep changes small and focused; match the existing stdlib-only style unless a dependency is strongly justified.
3. Open a pull request describing the motivation and include a quick before/after example where relevant.

Issues and feature requests are equally welcome — file them on GitHub.

License

Released under the MIT License. By contributing, you agree that your contributions will be licensed under the same terms.