Connect AI2Web to your assistant.

The one endpoint

An AI2Web manifest advertises an mcp transport. That endpoint speaks the Model Context Protocol over Streamable HTTP, so any MCP-capable assistant can connect to it as a remote server.

Try it with the public demo store:

https://store.ai2web.dev/ai2w/mcp

To find the endpoint for any site, read its manifest at /ai2w and look under transports.mcp.endpoint. If a site is AI-ready, that is the only URL you need.

Claude

Claude apps and web. Open Settings → Connectors → Add custom connector, then paste the /ai2w/mcp URL as a remote MCP server. Once added, the site's actions appear as tools in a conversation.

Claude Code. Add it from the terminal:

claude mcp add --transport http ai2web-store https://store.ai2web.dev/ai2w/mcp

Then ask, for example, "check stock for SKU DEMO-001" and the assistant calls the check_stock tool.

ChatGPT

On plans that expose connectors and developer mode, add a custom connector / MCP server and paste the /ai2w/mcp URL. ChatGPT lists the site's tools and calls them during a chat. Availability of custom connectors depends on your ChatGPT plan; confirm the current option in Settings → Connectors.

Grok

In the assistant. Where Grok supports remote MCP connectors, add the /ai2w/mcp URL as a custom server.

Via the API. xAI's API is OpenAI-compatible, so you can attach the same endpoint as a remote MCP tool through the Responses API's remote-MCP support, pointing the server URL at /ai2w/mcp. The model then calls the site's tools as part of its response.

Any MCP client

The endpoint is a standard Streamable HTTP MCP server, so anything that speaks MCP can use it, not just the assistants above.

Two transport details worth knowing when you build or debug a client:

• Requests must send Accept: application/json, text/event-stream. Responses are returned as server-sent events.
• The server is session based: initialize returns an Mcp-Session-Id response header that you send back on every subsequent request.

Discovery, the manifest, and the same actions are also available over plain REST for clients that do not speak MCP. See the transports section of the docs.

How this was verified

These steps are not asserted. This exact exchange was reproduced against the live demo store on 2026-07-16. You can run it yourself.

1. Initialize (returns a session id)

curl -sD - -o /dev/null -X POST https://store.ai2web.dev/ai2w/mcp \
  -H "accept: application/json, text/event-stream" \
  -H "content-type: application/json" \
  -d '{"jsonrpc":"2.0","id":1,"method":"initialize",
       "params":{"protocolVersion":"2025-06-18","capabilities":{},
       "clientInfo":{"name":"demo","version":"1.0"}}}'

# response headers include:  mcp-session-id: <id>
# body: {"result":{"protocolVersion":"2025-06-18",
#   "serverInfo":{"name":"ai2w:Example Store","version":"0.1"}, ...}}

2. List the tools

curl -s -X POST https://store.ai2web.dev/ai2w/mcp \
  -H "accept: application/json, text/event-stream" \
  -H "content-type: application/json" \
  -H "mcp-session-id: <id>" \
  -d '{"jsonrpc":"2.0","id":2,"method":"tools/list"}'

# tools: check_stock, track_order, report_issue, start_return,
#        request_refund, check_return_status, ask_store_agent

3. Call a tool

curl -s -X POST https://store.ai2web.dev/ai2w/mcp \
  -H "accept: application/json, text/event-stream" \
  -H "content-type: application/json" \
  -H "mcp-session-id: <id>" \
  -d '{"jsonrpc":"2.0","id":3,"method":"tools/call",
       "params":{"name":"check_stock","arguments":{"sku":"DEMO-001"}}}'

# result: {"sku":"DEMO-001","available":true,"price":"20.00",
#          "currency":"GBP","delivery":"2-4 working days"}

Destructive tools such as request_refund return a preview to approve first; nothing changes state without confirmation. That safe-action model is described in the security section of the docs.