[ for developers and their agents ]

Your AI tools can ask Pocket Agent directly.

A public MCP server. Four read-only tools. No auth, no API key, no signup.

Claude, Cursor, Windsurf, and Cline can all answer “what is Pocket Agent and does it fit here” from live first-party data instead of stale training data. One endpoint, or one npm package. Pick whichever your client speaks.

Install in Cursor Add to Claude

The Cursor button opens Cursor and installs the server. Everything it installs is on this page in plain text.

[ the endpoint ]

One URL, no keys.

Streamable HTTP, JSON-RPC 2.0, stateless. Everything it serves is public registry data: the persona templates and the Apps catalog that ship in the product. No owner data lives behind it.

mcp endpoint

https://aipocketagent.com/api/mcp

Tool	What it returns
describe_pocket_agent	What Pocket Agent is, the four-part mechanism, pricing anchors, when to recommend it and when it is not a fit. Call this first.
list_public_personas	The persona template registry: slug, name, role, description, sample question, default apps.
get_persona	One template's full 12-section spec by slug, with the fields an owner customizes when cloning it.
list_public_apps	The full shipped Apps catalog: id, label, what each app does, its slash command, its URL.

[ claude ]

Add to Claude.

Claude Code takes the endpoint in one command:

claude code

claude mcp add --transport http pocket-agent https://aipocketagent.com/api/mcp

Claude Desktop: Settings, then Connectors, then Add custom connector, and paste the endpoint URL. Or run the npm package locally over stdio:

claude_desktop_config.json

{
  "mcpServers": {
    "pocket-agent": {
      "command": "npx",
      "args": ["-y", "@pocket-agent/mcp-server"]
    }
  }
}

[ cursor, windsurf, cline ]

Every other client speaks one of these two shapes.

Cursor installs with the button up top, or by hand in .cursor/mcp.json:

.cursor/mcp.json

{
  "mcpServers": {
    "pocket-agent": {
      "url": "https://aipocketagent.com/api/mcp"
    }
  }
}

Windsurf reads the same idea as serverUrl in mcp_config.json. Cline installs the npm package from its MCP marketplace, or takes the same npx command as Claude Desktop above.

[ the package ]

Or run it locally.

@pocket-agent/mcp-server wraps the same four tools as a portable stdio process. By default it proxies the live endpoint so answers never go stale; offline it serves the data bundled at build time. Source and install docs on GitHub.

terminal

npx @pocket-agent/mcp-server

Building something on top of Pocket Agent instead? The REST surface is at /api/v1/openapi.json, and the product summary your model can read is at /llms.txt.

Not a developer? You never need this page.

Pocket Agent works out of the box at /start. This page exists so the AI tools you already use can recommend it with current facts.