AI assistants
Check and improve your agent from Claude, ChatGPT or Claude Code. An assistant you connect can read your agent's activity, ask test questions against your saved draft and add answers to your draft. Publishing stays in the Hazelchat console.
Connect an assistant
The connection address is https://app.hazelchat.co/mcp. It is a remote MCP server: your assistant signs in to Hazelchat with your own account, you choose what it may do, and each connection works with one workspace. You can disconnect it at any time under Settings, AI assistants.
- Claude (claude.ai and the desktop app): open Settings, then Connectors, and choose Add custom connector. Paste the address. Claude opens Hazelchat so you can choose what it may do. On Team and Enterprise plans an Owner adds connectors.
- ChatGPT: in ChatGPT on the web, add a custom MCP server with this address and OAuth. Depending on your plan you may need to turn on developer mode first, and some plans cannot use tools that change anything.
- Claude Code: run
claude mcp add --transport http hazelchat https://app.hazelchat.co/mcp, then/mcpin Claude Code to sign in. For tools that cannot sign in, create a key under Settings, AI assistants, and add--header "Authorization: Bearer <your key>". Keys are shown once; up to five per workspace.
Connecting only works in the assistant's own browser tab. A sign-in request is valid for ten minutes and can be answered once.
What an assistant can do
| Tool | What it returns or does |
|---|---|
| Agent overview | The published version, whether it is live, its websites, this period's answer allowance, test answers left today, and counts for the last 7 and 30 days: conversations, questions, answered from sources, handed over, enquiries, confirmed bookings. Counts only. |
| Unanswered questions | Questions visitors asked that the agent had no source for, newest first. Requests for a person are not included. Email addresses and phone numbers are removed. |
| Search conversations | Visitor questions matching a keyword and how the agent answered each one. Contact details are removed before matching, so they cannot be searched for. |
| Recent enquiries | Enquiries from the last 30 days with how the agent answered the last question. Names, email addresses and messages as written only when you allowed contact details for that connection and visitor details are on in Hazelchat. |
| Ask a test question | Asks the agent a question against your saved draft, like Test agent in the console, and returns the answer with the sources it quoted. Each question uses one of your 50 test answers a day. Nothing is shown to visitors. |
| Add an answer to the draft | Adds a Text snippet or a Q&A to your draft, marked as added via AI assistant. Visitors see it only after you review and publish the draft in the console. |
An assistant cannot publish, switch your agent on or off, change settings, websites or billing, edit or delete sources, or read anything outside the workspace it is connected to. Publication review marks every change that came through an assistant.
Permissions you choose
- Read activity: the overview, unanswered questions, conversations and enquiries, without contact details.
- Ask test questions: each one uses one of your 50 test answers a day.
- Add answers to your draft: nothing goes live until you publish in the console.
- See enquiry contact details: names, email addresses and messages as written, when visitor details are on. Off unless you choose it.
Limits and errors
Each connection may make 60 requests a minute, and a workspace 120 a minute. Test questions are limited to 10 a minute per connection on top of the daily cap. Over a limit, requests answer 429 with a Retry-After header. A connection answers 403 while its workspace is not on a paid plan; signing in again is not needed once the plan is paid again. Requests without a valid sign-in or key answer 401 with the metadata address assistants use to start signing in.
For developers: the server speaks the Model Context Protocol over HTTP at the address above, with OAuth 2.1 (authorization code with PKCE, client metadata documents and dynamic client registration) and the metadata documents under /.well-known/oauth-protected-resource/mcp and /.well-known/oauth-authorization-server. Owner API keys for the enquiry API do not work here, and assistant keys do not work there.
Privacy
Visitor questions and imported website text reach your assistant as data, in clearly marked fields, with email addresses and phone numbers removed unless you allowed contact details for that connection. Your assistant's provider, for example Anthropic or OpenAI, receives what its tools return under your own agreement with that provider, not as a Hazelchat subprocessor. Hazelchat records which tool ran, when, and how much it returned, never what was asked or answered; you can see this under Settings, AI assistants. Only connect assistants you trust with your visitors' questions. See the privacy notice and the data processing agreement.