Browse Anything#

This skill lets you delegate any web task to a real browser driven by an autonomous AI agent. You give a natural-language prompt; BrowseAnything opens Chromium, navigates, clicks, types, solves CAPTCHAs, and returns the result — including a screenshot.

When to use#

Trigger this skill whenever the task requires the live web, e.g.:

• "Find the cheapest flight from X to Y next month"
• "Log into my Notion and pull the latest entries from this database"
• "Fill out this Google Form with the following answers"
• "Check whether is down right now"
• "Buy item Z if it's under $50"
• "Scrape the top 20 results for query Q from "
• "Take a screenshot of after clicking Accept"

Do not use it for tasks the model can answer from internal knowledge, or for tasks that have a dedicated MCP/API the user already configured (prefer the more specific tool when available).

One-time setup#

1. The user must have a BrowseAnything API key ( ba_live_...). Direct them to https://platform.browseanything.io → Settings → API Keys to create one.

2. They export it once: bash

   export BROWSEANYTHING_API_KEY=ba_live_...
   
   

3. (Optional self-host) Set BROWSEANYTHING_API_URL=https://your-host to point at a self-hosted engine. Default is the hosted platform.

If BROWSEANYTHING_API_KEY is missing the scripts exit 2 with a clear message — surface that to the user verbatim.

Default workflow (high-level)#

For 95% of requests use the one-shot browse.py script. It creates a task, polls until done, and prints the result.

bash

python3 {baseDir}/scripts/browse.py "Find the cheapest direct flight from CDG to NRT in May, return airline + price + booking URL."

Useful flags:

• --model <name>: override the LLM (e.g. gpt-5.2, kimi-k2.6)
• --max-steps <n>: cap agent steps (default 80)
• --proxy <region>: e.g. us, eu
• --metadata '{"key":"value"}': attach JSON metadata
• --timeout <seconds>: max wait (default 900)
• --json: emit the full task object instead of a friendly summary

Exit codes:

CodeMeaning0Task completed successfully1Task failed (read stderr / error_message)2Auth/usage problem (missing key, insufficient credits, bad input)3Network unreachable4Local timeout (task may still be running on server)5Task is paused waiting for human input — see below

Low-level workflow (manual control)#

Use these when you need to fire-and-forget, run many tasks in parallel, fetch screenshots mid-execution, or react to requires_input.

bash

ID=$(python3 {baseDir}/scripts/create_task.py "Prompt...")
python3 {baseDir}/scripts/get_task.py "$ID" --field status
python3 {baseDir}/scripts/get_task.py "$ID"                # full JSON
python3 {baseDir}/scripts/get_screenshot.py "$ID" --out latest.png
python3 {baseDir}/scripts/list_tasks.py --limit 20
python3 {baseDir}/scripts/cancel_task.py "$ID"
python3 {baseDir}/scripts/status.py                        # backend capacity

Handling human-in-the-loop#

If a task can't proceed without information only the user has (a 2FA code, a clarification, a confirmation), it transitions to status requires_input. The high-level browse.py exits with code 5 and prints the question. To answer:

bash

python3 {baseDir}/scripts/submit_input.py <task_id> "the user's answer"

Then resume polling with get_task.py (or call browse.py flow again on the same id by polling manually). Always ask the user before inventing an answer for a requires_input prompt.

Authoring great prompts#

The agent works best with prompts that are concrete and verifiable.

• ✅ "On amazon.fr, search 'Sony WH-1000XM5', open the cheapest new listing shipped from Amazon, return seller + price + ETA."
• ❌ "find me good headphones"

Tips:

• Name the website explicitly when you know it
• State the success criterion ("return X, Y, Z")
• Mention any login state ("I'm already logged in, my session is in the saved profile") — though credentials should never be passed in plain text; prefer pre-saved sessions in the BrowseAnything dashboard
• Cap scope: one task, one outcome

Cost & limits#

• Tasks consume credits; tier-dependent step/concurrency caps apply
• Default per-task hard cap: 80 steps, 20 minutes
• Rate limit: 100 API requests/min/key
• Supported models include gpt-5.2, gpt-5.4, kimi-k2.6, anthropic/claude-haiku-4.5, gemini-3-flash-preview, gpt-4.1, llama-4, openai/gpt-oss-120b, plus mini variants. The available set depends on your tier; unsupported values return a hard error rather than falling back. Copy the exact string from the API error message when retrying.

Pitfalls & troubleshooting#

• Model names are exact strings. The API validates the --model value strictly (e.g. gpt-5.2 works, gpt5.4 without a hyphen does not). If you get Invalid model, retry with the exact name from the API error message.
• Cancel only works on running tasks. cancel_task.py returns Task not found or cannot be cancelled for tasks that have already failed or completed. Check status with get_task.py --field status first.
• Human-in-the-loop blocks billing. A task stuck on requires_input consumes concurrency but not steps; answer promptly or cancel to free the slot.
• Foreground timeouts may be clamped by the host environment. If the terminal tool rejects a 900 s wait, run browse.py in the background ( background=true, notify_on_complete=true) and poll with get_task.py until it finishes.
• Inspect requires_input messages before replying. The agent sometimes embeds the completed answer inside its question (e.g. a table of flight results). If the task is effectively done, cancel it rather than submitting unnecessary input.

More#

• REFERENCE.md — full API surface, request/response shapes, status enum
• EXAMPLES.md — copy-paste prompt patterns for common scenarios
• README.md — install instructions for Claude Code, OpenClaw, Cursor, Codex, Gemini, Windsurf
• references/recurring-scraping-pipeline.md — architecture for daily automated scraping, deduplication, enrichment, and dashboard reporting (real estate, price monitoring, job boards, etc.)