DoConure

How do I let an AI agent write the integration?

This API is meant to be integrated by Claude Code, Codex, Cursor or whatever else you have open. Two files exist for that, and a button that writes the prompt for you.

The two files

FileWhat is in it
/llms.txtA short index: what this service is, and one line per documentation page. Cheap to fetch and enough to decide what to read next.
/llms-full.txtEverything, in one fetch: every endpoint, every setting, the response envelope, every outcome code with its number and whether it is retryable, then the whole guide as text. About 100 KB.

Both are generated from the running service. The settings come out of the request model, the outcome numbers out of the error catalogue, the response shape out of the code that builds it. They cannot describe a version of the API that does not exist, which is the whole reason to point an agent at them rather than let it infer the contract from an example.

The button

Your console has an Integrate with AI button on each template. It writes a prompt that already names your template id, the fields that template reads, and a payload of the right shape — the three things an agent cannot look up — and points it at /llms-full.txt for everything it can.

The prompt never contains your API key, on purpose. It says to read the key from the DOCONURE_API_KEY environment variable. A prompt is pasted into a chat transcript, stored by whoever runs the model, and often ends up in a file that gets committed; a key that reaches all three has been published. Reading it from the environment is also the integration you would want anyway, so there is nothing to trade off.

Rules for an agent

If you are writing the prompt yourself, these are the five that matter. They are the ones that are cheap to state and expensive to get wrong.

  1. Read /llms-full.txt before writing any code. Do not infer the request shape from an example. Unknown keys anywhere in the body outside data are refused, not ignored, so a plausible-looking invented setting fails the whole call (data is yours, and is free-form).
  2. Read the API key from the environment (DOCONURE_API_KEY). Never write it into a source file, a committed config, or a log line.
  3. Branch on code, never on message. Every failure carries a stable code and a four-digit number; the message is a hint for a human reading a log and may be reworded at any time.
  4. Retry only when the body says "retryable": true. A TEMPLATE_DATA_MISMATCH retried a hundred times fails a hundred times, and each attempt is billed as work done.
  5. Surface output.warnings. A call can succeed having done less than it was asked. TEMPLATE_FIELDS_MISSING means the PDF was produced with blanks in it and names the fields; swallowing that warning is how a customer finds out from their own customer.

Two more that are about this API in particular: generation is synchronous, so use a client timeout of at least 120 seconds; and send "responseFormat": "json" inside settings if you want the ids, the warnings and the document in one JSON body rather than raw PDF bytes with the identifiers in headers.

Model Context Protocol

An official MCP server is planned and not built yet, so that this page does not promise one. When it exists it will expose the operations an agent actually needs — list templates, fetch a template's expected fields, validate a payload before spending a credit on it, generate a document, read usage — and it will be announced here and in /llms.txt.

Until then, /llms-full.txt plus an agent that can make HTTP requests is the whole integration, and it is enough: there is one endpoint to learn and no SDK to install.

Next: First document · Outcomes and error codes · Generating a document