Skip to main content

​
Simple extraction (most common)

Get the page title. 
{
  "url": "https://example.com",
  "instruction": "Extract the page heading.",
  "output": {
    "type": "object",
    "properties": { "heading": { "type": "string" } },
    "required": ["heading"]
  }
}

​
Array extraction with count

Get the top 5 stories. 
{
  "url": "https://news.ycombinator.com",
  "instruction": "Extract the top 5 story titles and points.",
  "output": {
    "type": "object",
    "properties": {
      "stories": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": {
            "title": { "type": "string" },
            "points": { "type": "number" }
          },
          "required": ["title", "points"]
        }
      }
    },
    "required": ["stories"]
  }
}
The planner enforces counts extracted from the instruction (top 5, 3 methods, first ten). Over-extraction is truncated; under-extraction triggers a retry with the stronger fallback model.
Click into a category, then extract. 
{
  "url": "https://books.toscrape.com",
  "instruction": "Click the Science sidebar link, then extract the first 3 book titles and prices on that page.",
  "output": {
    "type": "object",
    "properties": {
      "books": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": {
            "title": { "type": "string" },
            "price": { "type": "string" }
          },
          "required": ["title", "price"]
        }
      }
    },
    "required": ["books"]
  }
}
The planner verifies the URL actually changed before extracting.

​
Form fill + submit

Submit a contact form. 
{
  "url": "https://example.com/contact",
  "instruction": "Fill the form and click submit.",
  "input": {
    "name": "Alice Nguyen",
    "email": "alice@example.com",
    "company": "Example Co",
    "message": "Hi, interested in a demo."
  },
  "output": {
    "type": "object",
    "properties": { "submitted": { "type": "boolean" } },
    "required": ["submitted"]
  }
}
Field names in input are matched against visible form labels + name attributes. Missing fields are left blank rather than filled with fabricated data.

​
Login + scrape

Log in, then extract from the post-login page. 
{
  "url": "https://quotes.toscrape.com/login",
  "instruction": "Log in with the given credentials, then go to the homepage and extract the first 3 quote texts.",
  "input": { "username": "user", "password": "user" },
  "output": {
    "type": "object",
    "properties": {
      "quotes": {
        "type": "array",
        "items": {
          "type": "object",
          "properties": { "text": { "type": "string" } }
        }
      }
    }
  }
}
Coming soon: profile + saveProfile fields so you can persist the logged-in session and skip the login step on subsequent tasks. See the session-profiles plan for status.

​
Search, click first result, extract

Search Wikipedia, extract from the result page. 
{
  "url": "https://en.wikipedia.org/wiki/Main_Page",
  "instruction": "Search for \"Rust (programming language)\", click the top result, extract the year it was designed.",
  "output": {
    "type": "object",
    "properties": { "designedYear": { "type": "number" } },
    "required": ["designedYear"]
  }
}

​
Action-only (no data extraction)

Just submit a form; no response data needed. 
{
  "url": "https://example.com/form",
  "instruction": "Fill the form using the input and click submit.",
  "input": { "name": "Alice", "email": "alice@example.com" }
}
Leave output unset; you get back status: SUCCESS and no data.

​
CAPTCHA handling

CAPTCHAs are auto-solved when visible — reCAPTCHA v2 and Cloudflare Turnstile are supported via CapSolver. If a solve fails, the task returns status: CAPTCHA_FAILED and your code can retry on a fresh sandbox.

​
Task-level timeout

{
  "url": "...",
  "instruction": "...",
  "timeoutMs": 120000
}
Default is 240_000 (4 min). Long multi-step flows should raise this; simple extractions can leave it at default.

​
Success condition

For flows where the planner can’t tell when “done” from the schema alone: 
{
  "successCondition": "The page shows a 'thank you' confirmation message."
}
The planner uses this to know when to stop acting and finalize.