API

Ask a website questions from your own code, and get back the answer plus the pages it came from.

1. Get a key

Sign in, then open Profile & settings → API & MCP and create one. The key is shown once: we store a one-way hash of it and cannot show it again. If you lose it, revoke it and create another.

Each key carries scopes, and can be restricted to particular websites. A key never widens what your account can do - it is a way to sign in, not extra authority.

Send it as a bearer token, or as X-API-Key if your client only offers a custom header:

Authorization: Bearer fsc_sk_...

2. Ask a question

curl -X POST https://firesitechat.com/api/v1/answer \
  -H "Authorization: Bearer $FSC_KEY" \
  -H "Content-Type: application/json" \
  --data-binary @- <<'JSON'
{"domain":"example.com","message":"What are your opening hours?"}
JSON

{
  "answer": "We're open 9am-5pm Monday to Friday...",
  "grounded": null,
  "engine": "lite",
  "model": "...",
  "sources": [{"url": "https://example.com/contact", "title": "Contact us"}],
  "usage": {"input": 1840, "output": 96, "total": 1936, "llm_calls": 1},
  "creditsSpentMc": 2104,
  "chatId": "api-9f2c..."
}

What the fields mean

Your history

API and MCP answers are saved to your account's chat history by default and appear in the dashboard sidebar, marked { } (API) or 🤖 (a connected assistant), so you can see what your integrations asked. Send "save": false for a one-shot that leaves no trace.

Verified answers

Add "engine":"agentic" to run the checking agent. It is slower and roughly three times the cost, and it is the only way grounded becomes true.

Streaming

POST /api/v1/answer/stream returns the same answer as Server-Sent Events, so the first words arrive in well under a second. Event types: start, notice, answer_delta, draft, verdict, answer, done, error.

3. Search and read the content yourself

curl -X POST https://firesitechat.com/api/v1/search \
  -H "Authorization: Bearer $FSC_KEY" -H "Content-Type: application/json" \
  --data-binary @- <<'JSON'
{"domain":"example.com","query":"opening hours","k":6}
JSON

Returns the passages with their source URLs and scores - no model in the loop. GET /api/v1/sites/{domain}/pages lists what we hold, and GET /api/v1/sites/{domain}/pages/{pageId} returns one page's text.

That page response reports chunkCount (what the page claims) and chunksHeld (what actually came back), so a half-failed indexing run shows as a mismatch rather than as an innocent-looking short page.

Which websites you can search and read

The two calls answer to different rules, on purpose.

• POST /api/v1/search works on any website in our index, not just yours. That is the point of it: you can ask about a business you do not own, the same way /api/v1/answer can. It is metered against your own credits.
• GET /api/v1/sites/{domain}/pages and .../pages/{pageId} are narrower. Together they walk every URL we hold for a website and hand back the full text of each, so they are limited to websites on your own account and websites listed in our public directory. Anything else answers 403 and points you at /api/v1/answer.

If you run a website and would rather it were not readable this way, the opt-out in your console removes it from the public directory, which closes the page-level calls to everyone but you. It does not close search: passages from an indexed website stay retrievable by any key. Content we index is content your website already serves publicly, and we crawl only what robots.txt allows.

4. Manage websites

Indexing runs in the background - poll GET /api/v1/sites/{domain} until status is ready. Removing a website does not delete its index: it is shared with everyone else asking about that domain, and re-adding would pay for the crawl again.

Errors

"Field required" for fields you did send

Your shell received typographic quotes (“ ”) instead of ", which happens when a command is copied out of a rendered document. The JSON still parses, but into a single field name, so the fields really are missing. On macOS, turn off System Settings → Keyboard → Text Input → Input Sources → Edit → Use smart quotes and dashes. The heredoc form used above is immune to the shell, and this check is definitive:

grep -q $'[‘’“”]' body.json && echo "BAD: curly quotes" || echo "OK"

Reference

The complete, always-current reference is generated from the server itself: interactive docs · OpenAPI document.

Trying calls from the browser: open the interactive docs, click Authorize (top right), paste your key, then Try it out → Execute on any endpoint. The key is remembered while the tab is open, and the panel shows you the exact curl for whatever you just ran.

Guide · MCP & ChatGPT · Get a key