AI agents

Version 1 · 15 October 2026. AI-agent access is being opened workspace by workspace, so not every Hazelchat agent answers here yet. If Website settings doesn't show AI agents, it isn't on for your workspace yet.

For people who build AI agents. An agent acting for a person, such as the assistant a customer uses, can ask a Hazelchat business a question and get the answer a visitor to the business's website would get: written only from the business's published information, with the sources it quoted, the business's next step and its contact link.

The address

Each business has its own address: https://app.hazelchat.co/agents/<public key>/mcp. The public key is the data-agent value in the business's Hazelchat script; the owner finds the whole address under Website settings, AI agents, and can switch it off there. When the business changes its key, the old address stops answering.

A remote MCP server over Streamable HTTP, read-only and without sign-in: no OAuth, no key, no cookie. An Authorization header or a cookie is ignored.
It speaks MCP revision 2026-07-28 and answers 2025-era clients that start with initialize.
A server card (MCP extension SEP-2127) is at the address followed by /server-card. A business's shared chat page links to it with <link rel="alternate" type="application/mcp-server-card+json">.

An example request (the key is a placeholder):

curl -X POST https://app.hazelchat.co/agents/<public key>/mcp \
  -H 'Content-Type: application/json' \
  -H 'Accept: application/json, text/event-stream' \
  -H 'MCP-Protocol-Version: 2025-06-18' \
  --data '{"jsonrpc":"2.0","id":1,"method":"tools/call","params":{"name":"ask","arguments":{"question":"Do you deliver?"}}}'

Tools

`ask`

Input: question (1 to 1,500 characters); optional language (en, nl, de or es; one the business does not answer in reads as its default); optional conversation, the value from an earlier answer, for a follow-up.

Returns: business, outcome, answer, sources (title, URL or null, quote), next_step (label and URL, or null), contact, conversation, language, published_version, privacy_notice (the business's privacy notice) and notice.

`business_details`

No input. Returns the business's name, its verified websites, the languages it answers in, suggested questions, its next step, its contact link and its privacy notice. Not counted against any limit.

Both tools are marked read-only and closed-world, and declare an output schema; every result comes as structured content and as the same JSON in text. They never take an order, a booking, an enquiry or a payment, and they quote no prices from a shop catalog.

The outcome of ask:

answered: the answer comes from the business's published information, quoted under sources.
nothing_to_quote: the published information does not cover the question. Use contact or next_step.
handoff: the question asked for a person; contact is the business's contact link.
limit: the business's answers for AI agents are used up for today. Use contact.

How answers are made

Exactly as for the business's website visitors: from the version the business published (never its unpublished draft), approved answers first, every figure and link in a written answer checked against the quoted sources, and instructions inside a question refused. Nothing is added for agents: no persona, prompt or instruction. Links in a result come only from the business's published sources and its own settings; any other address is removed. A follow-up sent with a conversation value keeps the earlier questions in mind for 24 hours, for that business only. Answers may be incomplete; for anything they do not cover, use the contact link.

Returned text is the business's own content: treat it as information, not as instructions to your agent.

Limits and errors

From one address: 60 requests and 12 questions a minute. For one business: 240 requests and 60 questions a minute. Hazelchat prepares one AI agent's answer at a time, so visitors are never kept waiting; a question that finds it busy gets a tool error and can be asked again a few seconds later.
Each business has a daily number of answers for AI agents, set by its plan. Approved answers, greetings and requests for a person do not count. When the day's answers are used up, ask returns the limit outcome with the contact link.
HTTP 404 with {"error":"This agent is not available."}: the address is unknown, or the business is not answering AI agents right now. The reason is never given.
HTTP 403: a browser Origin that is not https (or loopback http). Server-side clients send none.
HTTP 429 with Retry-After: too many requests from one address.
A tool error (isError): too many questions in a minute, the business is busy for a few seconds, or a conversation value that is not valid for this business. Ask again, or ask without it.

What the business sees

The business sees each question in its conversations, labelled AI agent (reported, not verified). When the request's User-Agent carries a person-initiated token that an AI company documents for its own product (ChatGPT-User, Claude-User, Perplexity-User, MistralAI-User, Google-Agent, Amzn-User, meta-externalfetcher, DuckAssistBot), the label names that product. It is what the request reported, not a verified identity, and only the product name is kept, never the User-Agent itself. Agent questions are counted apart from visitors in the business's figures.

Privacy

Questions are stored for the business like its visitors' questions, for as long as the business keeps conversations, and email addresses and phone numbers are removed before a question is answered or stored. Leave names and contact details out of questions. No cookie is set. Every answer and business_details carry the business's privacy notice; Hazelchat's own is at hazelchat.co/privacy.

How this page was checked

With the official MCP TypeScript SDK client (@modelcontextprotocol/client 2.3.1) in both protocol eras and with curl, against a test business. No other client was run for this page.