# AI agents

> Give LLMs and agents eyes on the web with screenshots, Markdown and vision, and point them at these docs.

## Pages as Markdown

`format=markdown` renders a page in a real browser and returns its content as Markdown: JavaScript-rendered content included, scripts and styles removed. It's a compact way to put a web page into an LLM's context.

```bash
curl -X POST "https://shotkit.net/api/take" \
  -H "X-Access-Key: YOUR_ACCESS_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url":"https://example.com","format":"markdown","block_cookie_banners":true,"selector":"main"}' \
  --fail-with-body -o shot.md
```

`selector` narrows it to the part that matters and saves tokens. For pages behind a login, pass [cookies or an Authorization header](https://shotkit.net/docs/options/request.md).

## Screenshots for vision models

Vision models read screenshots well. Keep them small: a 1280-wide viewport capture as JPG is plenty, and `image_width` scales it down further:

```bash
curl -X POST "https://shotkit.net/api/take" \
  -H "X-Access-Key: YOUR_ACCESS_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url":"https://example.com","format":"jpg","image_quality":70,"image_width":1024,"block_cookie_banners":true,"block_chats":true}' \
  --fail-with-body -o shot.jpg
```

To get the picture and the text in one render, add `metadata_content` with `response_type=json`; see [HTML and Markdown](https://shotkit.net/docs/html-and-markdown.md#a-screenshot-and-the-content-in-one-request).

## Ask a question about the page

With your OpenAI key and a `vision_prompt`, the screenshot is sent to an OpenAI vision model and its answer comes back with the result:

```bash
curl -X POST "https://shotkit.net/api/take" \
  -H "X-Access-Key: YOUR_ACCESS_KEY" \
  -H "Content-Type: application/json" \
  -d '{"url":"https://example.com","response_type":"json","openai_api_key":"sk-YOUR_OPENAI_KEY","vision_prompt":"Is there a cookie banner on this page? Answer yes or no.","vision_max_tokens":10}' \
  --fail-with-body -o shot.json
```

```json
{
  "screenshot_url": "https://shotkit.net/cdn/files/…/….jpg",
  "vision": { "completion": "No" }
}
```

Your key is used for this request only and never stored. OpenAI errors fail the request with `vision_error`.

## Give your agent these docs

Every page of this documentation is plain Markdown for agents:

- Add `.md` to any docs URL: [https://shotkit.net/docs/quickstart.md](https://shotkit.net/docs/quickstart.md).
- Or request the normal URL with `Accept: text/markdown`.
- [https://shotkit.net/llms.txt](https://shotkit.net/llms.txt) lists every page with a summary ([llms.txt convention](https://llmstxt.org)).
- [https://shotkit.net/llms-full.txt](https://shotkit.net/llms-full.txt) is the whole documentation in one file.

In the Markdown version, examples are cURL commands, so an agent can run them directly. The **Copy page** button at the top of each page copies the same Markdown, ready to paste into a chat.

A system prompt line that works well:

```text
To capture web pages, use the shotkit API. Its documentation is at https://shotkit.net/llms.txt. The access key is in the SHOTKIT_ACCESS_KEY environment variable.
```