AgentRank

MCP server for AI agents to search real, source-backed businesses. Every result carries a source URL and a trust score, so your agent acts on facts — never invented phone numbers or addresses. Coverage: Brazil, US, Europe (~2.7M businesses, growing).

Works with any MCP client (Claude Desktop, Cursor, …). Public beta · free key, no invite · website: https://agentsafepath.com · official MCP registry: io.github.saviodamato/agentrank-mcp.

Use cases

• Business search / company lookup with source-backed data.
• Local business search for agents (city + category + contact).
• Company profile: contacts, address, website, provenance.
• Source-backed company data instead of hallucinated details.

Tools

• search_businesses(intent, limit=10) — find businesses by a natural-language intent (e.g. "farmácia em Copacabana com telefone", "software company in California"). Returns ranked results with fit_score, trust_score, justification, risks and the source URL of each fact.
• get_business_profile(company_id) — full profile of a result: location, contact channels, hours, offers, policies and evidence, each with provenance and source.

1. Get an API key

Register your agent (free, no invite) at https://agentsafepath.com/registrar.html — you receive a key once. It identifies your agent and is sent as Authorization: Bearer <key>.

2. Add it to your MCP client

Claude Desktop

Edit claude_desktop_config.json (Settings → Developer → Edit Config) and add:

{
  "mcpServers": {
    "agentrank": {
      "command": "uvx",
      "args": ["agentrank-mcp"],
      "env": { "AGENTRANK_API_KEY": "ak_your_key_here" }
    }
  }
}

Restart Claude Desktop. You'll see search_businesses and get_business_profile available. (uvx comes with uv; or use the "From source" option below.)

Cursor

Add the same block under mcpServers in your Cursor MCP settings.

From source (no packaging)

pip install -r requirements.txt
AGENTRANK_API_KEY=ak_your_key_here python -m agentrank_mcp.server

And point your client's command at python with args ["-m", "agentrank_mcp.server"] and the folder on PYTHONPATH.

Environment variables

The key is never logged.

Beta & limits

Public beta. Free tier: ~100 queries/day, 1500/month per agent (HTTP 429 with Retry-After when exceeded). One key per agent; the key is shown once at registration.

How agents should use the data

Every result carries sources (source URLs). Verify provenance before acting — data without a source should not be used. trust_score (0–100) reflects how safe the entity is to commit to; fit_score reflects match to the intent.

mcp-name: io.github.saviodamato/agentrank-mcp