agent

Default browser automation agent — click, fill forms, navigate, log in, and extract structured data from any website using a natural-language goal, or run the same task across multiple sites in parall

SKILL.md

GitHub

TinyFish Agent

Browser automation via the bundled TinyFish MCP server. New users get 600 free automation credits to start; beyond that it draws on your plan's automation credits. Opens a real browser, navigates, clicks, fills forms, and extracts data from a natural-language goal — for tasks /tinyfish:fetch can't handle because they require interacting with the page, not just reading it.

run_web_automation

  • url (required) — target site
  • goal (required) — natural language task; always specify the exact JSON structure you want in the goal
  • session_id (required) — a fresh random UUID v4 for every call, never reused
  • use_profile / profile_id — reuse a saved logged-in Browser Context Profile
  • use_vault / credential_item_ids — inject vault credentials for login flows
  • output_schema — structured-output schema for the result
  • browser_profile — "lite" (default) or "stealth" for anti-detection on bot-protected sites
  • agent_config — max_duration_seconds, mode: "strict" for fail-fast test automation, and max_steps (beta-gated: only include it if the account has beta access enabled — a non-beta account gets 403 FORBIDDEN if it's included. Omit it to use the default of 150.)
run_web_automation(
  url="https://example.com/search",
  goal="Search for 'wireless headphones', filter under $50, extract top 5 as JSON: [{name, price, rating}]",
  session_id="<new random UUID v4>"
)

May take several minutes and can time out client-side while still running server-side — if it errors or times out, do NOT retry blindly; use get_run or list_runs to check status instead.

Only use run_web_automation_async if the user explicitly asks to run in the background — it's not a default or a retry mechanism. Poll with get_run every 30-60s.

Multiple independent sites — use batch_create, not repeated calls.

batch_create / batch_status / batch_cancel

For the same task across 2+ URLs:

batch_create(runs=[
  {url: "https://pizzahut.com", goal: "Extract pizza prices as JSON: [{name, price}]"},
  {url: "https://dominos.com", goal: "Extract pizza prices as JSON: [{name, price}]"}
])

Up to 8 runs per batch, returns all run IDs immediately. batch_status(run_ids) polls every 30-60s until all reach a terminal state. batch_cancel(run_ids) stops running/pending runs.

Managing runs

  • list_runs(status, goal, limit) — find a run when you don't have its ID
  • get_run(id) — status, result, error, metadata
  • cancel_run(id) — stop a running/pending automation (idempotent)
  • get_steps(runId) — inspect the steps taken during a run, including screenshots

create_browser_session / list_browser_sessions

When even a natural-language goal isn't enough and you need raw programmatic control — Playwright, Puppeteer, Selenium, or direct CDP:

create_browser_session(url="https://example.com")
# Returns: session_id, cdp_url (wss://...), base_url

list_browser_sessions reviews active or past sessions.

Notes

  • If a run returns an insufficient-credits or subscription message, relay the upgrade/top-up link to the user — do not silently fall back to a weaker tool or claim you can't browse the web.
  • Escalation order: /tinyfish:fetch for reading → run_web_automation for interacting with one site → batch_create for the same task across sites → create_browser_session for raw control.

$ARGUMENTS