Authentication
This endpoint is public. It takes no credential and no organisation — it is what our own marketing site and AI answer engines read.
This endpoint takes no organisation id. Your key already identifies the organisation it belongs to, and the response is scoped to it.
Try it
Replace anything in angle brackets with your own values, and the key placeholder with a key from your dashboard.
curl -X POST https://api.zinndigital.com/v1/sites/{siteId}/ide-chat \
-H "Content-Type: application/json" \
-d '{ "messages": <IdeChatMessage[]> }'Signed in? The API console in your dashboard fills in your real organisation id and your own key, and runs the request against the live API so you can see the actual response. Open this endpoint in the API console
Details
The conversation panel inside the site's web editor. Unlike `askSiteIdeAssistant` — which answers one question about one file — this carries a transcript and a catalogue of **tools** the editor is willing to run, so the assistant can answer questions about the customer's own hosting rather than only about their code. Authenticated by the same signed, site-bound grant as the rest of the editor's endpoints, and metered to the same organisation on the same `web_ide` surface. The model is the organisation's own choice, or the cheapest priced one when they have expressed none — Claude, ChatGPT, Gemini or Grok, funded from Zinn® credit or the customer's own provider key. **The engine never runs a tool.** When the model asks for one, the response carries `tool_calls` and an empty `answer`, and the conversation is unfinished: the editor invokes them through its own MCP client — where the customer's confirmation happens for anything destructive, and where the MCP server's write gate and spend caps apply — and posts the results back on the next turn. A response with an empty `answer` and no `tool_calls` is never returned; that is a `422` carrying its own sentence, because the call was paid for either way and a blank panel is indistinguishable from a broken one. The transcript is authored inside the customer's own jail, so it is treated as data: roles are mapped from a closed set, the tool list is bounded and coerced, and a malformed tool is dropped rather than offered.
Parameters
| Name | Type | Required | What it is |
|---|---|---|---|
siteId (path) | Uuid | Yes | Site ID (UUIDv7). |
Request body
| Name | Type | Required | What it is |
|---|---|---|---|
messages | IdeChatMessage[] | Yes | The conversation so far. The engine is stateless, so the whole transcript arrives on every call and the caller bounds it. |
tools | IdeChatToolOffer[] | No | What the editor can run — its MCP servers' tools and any the customer's own extensions contribute. Bounded and coerced on arrival; a malformed entry is dropped rather than offer… |
path | string | No | The file the customer is looking at, if any. |
context | string | No | That file's contents, truncated to the assistant's context cap. |
model | string | No | A model from `/ide-assist/models`. Omit for the automatic choice, which is the cheapest priced model the organisation can run. |
Response
| Name | Type | Required | What it is |
|---|---|---|---|
answer | string | Yes | Empty when the assistant asked for tools instead. Never empty alongside an empty `tool_calls` — that is a `422` with its own sentence, because the call was paid for either way. |
tool_calls | IdeChatToolCall[] | Yes | — |
model | string | Yes | — |
funding | string | Yes | Which purse ran it — the same value the AI page renders, from the same decision. The editor shows "your Zinn® credit" or "your own API key" beneath the answer. |
turns | integer | Yes | How many turns of this conversation have already asked for tools, so the editor can stop a model that will not settle rather than paying for it indefinitely. |
Errors this endpoint can return
404 · 422 · 429