@ai2web/connector

The agent side of AI2Web - an MCP server you load into Claude (custom connector) or ChatGPT (App via Apps SDK). It fronts the Discovery Network so an assistant can find AI-ready sites and act on them, with no vendor needing to adopt ai2w - only to allow third-party MCP connectors (Claude and ChatGPT already do).

This is the "we own both sides" piece that defeats the cold-start problem.

Tools it gives the assistant

| Tool | Does | |---|---| | find_sites | Search the Discovery Network by capability / type / text. | | describe_site | Fetch a site's manifest + negotiate the capabilities/tools available to you. | | call_site_action | Call a site's declared action. Approval-gated actions return a preview, never auto-execute. |

Use as an MCP server

The package ships a runnable stdio MCP server (bin ai2web-connector). It needs @modelcontextprotocol/sdk installed and an AI2WEB_DIRECTORY pointing at a deployed Discovery Network.

Run it directly:

AI2WEB_DIRECTORY=https://directory.ai2web.dev npx @ai2web/connector

Add it to Claude Desktop (claude_desktop_config.json):

{
  "mcpServers": {
    "ai2web": {
      "command": "npx",
      "args": ["-y", "@ai2web/connector"],
      "env": { "AI2WEB_DIRECTORY": "https://directory.ai2web.dev" }
    }
  }
}

In Claude.ai / ChatGPT you can instead add the hosted remote connector (a Cloudflare Worker exposing /mcp) as a custom connector; see the ai2web-cloud repo.

Configuration

| Env | Purpose | Default | |---|---|---| | AI2WEB_DIRECTORY | URL of the Discovery Network the connector queries | https://directory.ai2web.dev |

Tools

Once connected, the assistant gets three tools: find_sites, describe_site, and call_site_action (approval-gated actions return a preview instead of executing).

Verified

The full discover → describe → negotiate → act loop (with approval gating) is proven in-process by ai2web-js/scripts/spike.ts. The stdio entrypoint (src/stdio.ts) starts this server over stdio for local MCP clients.