The index, as an MCP server

The server is remote and uses the Streamable HTTP transport — no install, no API key. Point a compatible MCP client at:

https://agentsearchengine.app/api/mcp

Claude — this is a remote server, so it is added through the connector UI rather than a config file. On an individual Pro or Max plan: Customize › Connectors, then + › Add custom connector, and paste the URL above. On Team or Enterprise an Owner adds it under Organization settings › Connectors › Add › Custom › Web. Custom connectors are available on Free, Pro, Max, Team and Enterprise, and free accounts are limited to one. Note that claude_desktop_config.json is for local servers only and will not work for this one.

Cursor — add it to ~/.cursor/mcp.json:

{
  "mcpServers": {
    "agent-search-engine": {
      "url": "https://agentsearchengine.app/api/mcp"
    }
  }
}

Claude Code —

claude mcp add --transport http agent-search-engine https://agentsearchengine.app/api/mcp

ChatGPT — custom MCP connectors sit behind developer mode, which OpenAI documents as available to Pro, Plus, Business, Enterprise and Education accounts on the web, enabled under Settings › Security and login. OpenAI flags developer mode as elevated risk; this server is read-only, but that judgement is yours to make.

Anything else — point a client at the URL, or check it first with the MCP Inspector (npx @modelcontextprotocol/inspector), choosing the Streamable HTTP transport.

Client setup paths and plan tiers are each vendor's to change, and were last checked against their own documentation on 28 September 2026. We have verified this server against the MCP SDK and the official MCP client library; we have not interactively tested every client listed above, so treat these as directions rather than a compatibility guarantee.

This server implements Streamable HTTP at protocol revisions 2024-10-07 through 2025-11-25. The current published revision is 2026-07-28, which this server does not yet implement, so it is worth being precise about what that means rather than claiming every client works.

• Clients that negotiate normally: supported. The official 2026-07-28-era client SDK defaults to the legacy handshake and its automatic mode falls back to it, so it connects here and can call all four tools. That is how today's Claude, Cursor and Inspector builds reach the server.
• Clients pinned to 2026-07-28: not supported. A client configured to require that exact revision will refuse to connect, by its own design — it will not fall back. Sending MCP-Protocol-Version: 2026-07-28 on a request returns HTTP 400.
• POST only. The endpoint accepts POST; GET, HEAD and DELETE return 405. There are no protocol-level sessions, and the newer server/discover method returns “method not found”, which is the signal a modern client uses to fall back.

• Organic, never sold. The MCP surface is not for sale — no sponsored placements, ever, by construction. Sponsorship on the website is separate, labeled ad inventory that never touches these results.
• It is designed to abstain rather than guess. When no candidate clears its confidence threshold, the search returns an empty result flagged abstained, with category suggestions instead of a fabricated recommendation. It is a threshold, not a guarantee: on our judged set of 97 deliberately out-of-scope queries it abstained on 92. Nothing is ever invented — but a weak near-match can still surface, which is why every result carries a link you can check.
• Every result is cited. Each record carries a canonical link back to its listing, so assistants that use us can link to us. Freshness timestamps travel with the response.
• Third-party data, labeled. Descriptions are curated from each tool's own public sources and returned as sanitized plain-text data — not instructions.

The ranking method is the same one documented on the methodology page. The full corpus is also available as a plain-text llms.txt feed.

The server answers from a bundled snapshot rather than querying the database on every call, so results are fast and stay available when the database is slow. Every response carries the two dates below, so an assistant can tell you how current the answer is.

• Index generated: October 5, 2026 — 454 records, 11 categories.
• Stars and activity observed: September 28, 2026.

• Read-only. Four tools, no writes, no accounts, no API key.
• What happens to your queries. The application code does not intentionally log or persist the text of your queries or the bodies of your requests: it holds them in memory to answer the call and keeps nothing. There is no analytics or telemetry in this endpoint at all. Separately, and like any website, Vercel processes standard request metadata — IP address, user agent, timestamp, path — as our hosting provider, under its own retention. We do not copy that into an application store, join it to query content, or build a profile from it.
• Only what is indexed. Asking for a tool the catalogue does not carry will usually produce an explicit empty result rather than a substitute, and a direct lookup by slug always answers found: false. Absence from the index is a limit of our coverage, not a judgement about the tool.
• Snapshot, not live. A listing added or corrected today appears after the next snapshot build, not immediately.
• Found something wrong? Tell us — corrections to a record are free and are never tied to sponsorship.