AgentSearch

The agent that helps agents find other agents.

agentsearch.luthersystems.com — a daily-scored, curated index of AI agents and MCP servers. Searchable, browseable, and callable by other agents.

This repository hosts the open-source agent surfaces for AgentSearch:

• mcp/ — Model Context Protocol server. Plug into Claude Desktop, Claude Code, Cursor, or any MCP client and call AgentSearch as a tool.
• agent-card/ — A2A agent card (agent.json) for the a2aregistry.
• openapi/ — OpenAPI 3.1 spec for the public HTTP endpoints.
• docs/ — Methodology, scoring rubric, and integration guides.

The scoring pipeline that powers the index lives in a separate repo (luthersystems/agent-discovery).

What does AgentSearch do?

Given a natural-language query like "read invoice pdf" it returns a ranked list of agents and MCP servers that can actually do that. Each result has:

• A search-match score (query_fit, judged by claude-haiku-4-5)
• An agent-quality score (live HTTP reachability × weighted Fame / Usability / Functionality / Call-Graph)
• A combined final_score = query_fit × agent_quality

It also runs a live Google search for the same query and shows which Google results are real agents (extracted + scored on the fly).

Calling AgentSearch from your agent

Via MCP (recommended)

npm install @luthersystems/agentsearch
# or:  npx @luthersystems/agentsearch

Add to your MCP client config (Claude Desktop, Claude Code, Cursor, …):

{
  "mcpServers": {
    "agentsearch": { "command": "npx", "args": ["-y", "@luthersystems/agentsearch"] }
  }
}

Tools exposed:

• search(query: string, limit?: number) — ranked agents for a query
• agent_details(name: string) — full scored profile of one indexed agent
• found_agent(url: string, name?: string, query?: string) — live-score an arbitrary agent URL using the same rubric (works for Google-found agents)
• browse(page?: number, page_size?: number) — paginated list of all indexed agents, ranked by overall quality
• stats() — index summary (totals, top sources/licenses, capability clusters)

Via HTTP

All endpoints are public, no auth required. See openapi/openapi.json.

curl -X POST https://agentsearch.luthersystems.com/api/search \
  -H "content-type: application/json" \
  -d '{"query": "read invoice pdf"}'

Via A2A

The agent card is at:

https://agentsearch.luthersystems.com/.well-known/agent.json

Submit URL to a2aregistry.org or load directly in any A2A-compliant client.

Releasing the MCP server

The mcp/ server is published to npm as @luthersystems/agentsearch.

Releases are fully automated — there is no manual npm publish and no npm token stored anywhere:

1. Bump version in mcp/package.json (optional — the workflow also derives the version from the tag).
2. Push a v* tag, e.g. git tag v0.1.1 && git push origin v0.1.1.
3. .github/workflows/npm-publish.yml builds the package and publishes it to npm.

Authentication uses npm trusted publishing over GitHub OIDC: the workflow requests a short-lived GitHub identity token, and npm verifies it against the trusted-publisher registered for this repo + workflow. Every release also carries a signed provenance attestation linking the published package to the exact commit and workflow run.

Tag form	npm dist-tag	Result
v1.2.3	latest	Default install — npx @luthersystems/agentsearch
v1.2.3-beta.1	next	Prerelease — npx @luthersystems/agentsearch@next

License

MIT — see LICENSE.