Coding agents

Upscrape publishes machine-readable context so a coding agent can implement an integration without guessing schemas or response behavior.

Give the agent authoritative inputs

For one platform, use its public LLM brief:

https://upscrape.com/scrapers/PLATFORM_ID/llm.md

For the complete public catalog, use:

https://upscrape.com/scrapers/llm.md

For the documentation index and complete product guide, use:

https://upscrape.com/llms.txt
https://upscrape.com/llms-full.txt

Every documentation page has a Markdown URL: append .md to a guide (/docs/quickstart.md) or /index.md to a platform or capability page (/docs/platforms/<platform>/index.md). Each HTML page names its Markdown URL in a <link rel="alternate" type="text/markdown"> tag and a Link header, and llms.txt lists them all. A platform page requested with Accept: text/markdown redirects to that URL, so fetch it directly and skip the hop. Prefer Markdown over parsing rendered HTML.

The per-platform brief includes the capability IDs, exact request contract, examples, input schemas, execution flow, and the current public OpenAPI document.

Keep secrets outside the prompt

Tell the agent to read the API key from UPSCRAPE_API_KEY. Never paste a real key into the conversation or generated source.

The resulting application should:

• send the bearer token from a server-side environment;
• use Prefer: wait only as a bounded optimization;
• handle both HTTP 200 terminal representations and HTTP 202 pending jobs;
• inspect state and success on job responses;
• use an idempotency key for logical application operations;
• validate capability input before sending it;
• treat raw result data as open-ended and untrusted.

Existing prompt generator

A platform-, capability-, and stack-aware integration prompt is published, without authentication, at /scrapers/{platform_id}/capabilities/{capability_id}/llm.md. Add ?stack=python, ?stack=node or ?stack=curl to pick the stack. The authenticated console exposes the same generator.

Review generated code

Before deploying agent-written integration code, verify the exact capability ID, example input, poll termination, retry deadline, error-code handling, and secret storage. Generated code should not invent a stable output model where the capability publishes open-ended JSON.