Browser Use AI Task Runner

Beta version 0.1.7. Runs a natural-language task on public websites and returns one final JSON result, with optional sanitized step rows.

Beta status

Start with short public-web tasks. The default Gemini 3.5 Flash-Lite passed live heading and two-page author extraction; the two-page case was repeated successfully with default fields omitted. Gemini 3.8 Flash remains selectable but returned provider HTTP 503 twice during final validation. Free-tier quota can also return HTTP 429. There is no guarantee that every task will complete. OpenAI and Anthropic adapters are available but have not been live-benchmarked for this release. Long responses, authenticated workflows, and broad adversarial coverage remain unverified.

Quick start

Supply a public start URL, its exact allowed domain, your provider/model, and a provider key in the secret API key field. Never put credentials inside the task.

{
  "task": "Read the page heading and return it as JSON.",
  "startUrls": ["https://example.com"],
  "allowedDomains": ["example.com"],
  "provider": "google",
  "model": "gemini-3.5-flash-lite",
  "maxSteps": 5,
  "outputSchema": {
    "type": "object",
    "properties": {"heading": {"type": "string"}},
    "required": ["heading"]
  }
}

The example does not contain a key. Alternative secret environment variables: OPENAI_API_KEY, ANTHROPIC_API_KEY, GOOGLE_API_KEY (or GEMINI_API_KEY). Provider key usage is billed by the provider separately. Provider wrappers exist for all three; Google gemini-3.5-flash-lite passed live heading extraction and the two-page author acceptance task. OpenAI/Anthropic benchmarks remain pending credentials.

What to use it for

• Extract public information into a specified JSON schema.
• Compare details on public product pages.
• Research specified company website fields.
• Navigate a permitted public workflow and summarize the outcome.

AI actions are not deterministic. Use the Playwright Browser Automation Runner for repeatable locator/action sequences.

Inputs and enforced limits

See the Input tab for the complete contract. Exact domain hosts only, no wildcard matching; include required asset hosts. Public HTTP(S) on 80/443 only.

• maxSteps: default 20, range 1-40; one agent action per step.
• maxTaskDurationSeconds: default 180, range 10-600; leave extra platform time for startup/cleanup.
• maxBrowserPages: default 3, range 1-5. Excess tabs are closed and execution is stopped.
• maxArtifactBytes: default 1 MB, maximum 2 MB, screenshot only.
• maxOutputBytes: default 100 KB, maximum 200 KB.
• Memory watchdog: 1536 MiB process tree; deploy with 2048 MiB.
• maxUserChargeUsd: default 0.20, maximum 1.00; Actor event limit, not a guarantee on BYOK or private-run compute.
• outputSchema: optional Draft 2020-12 JSON schema up to 20 KB; references and regex constraints are rejected. Validation is performed after redaction. Invalid results fail cleanly.
• screenshotMode: off by default, optional failure/completion/both. Visible sensitive forms suppress captures.
• includeStepTrace: false by default; only action names/step numbers/status, no raw parameters or model reasoning.
• proxyConfiguration: optional HTTP CONNECT upstream. Final destination is resolved locally and pinned to a public IP; compatibility with Apify Proxy remains a release gate.

Output

Canonical result is in KVS record FINAL-RESULT. Dataset has one final row; optional trace rows have recordType=step. Final fields: success/status, actorVersion/timestamp/inputHash, durationMs, provider/model, stepsUsed/completedSteps/chargedSteps, finalUrl, data/summary, screenshotUrl, warnings, errorCode/errorSummary. Input hash covers non-secret execution settings and start hosts, not the task, schema, queries or credentials. Final URL omits query/fragment.

Example from a real local Chromium + Browser Use integration using a scripted model (not a live-provider benchmark):

{"success":true,"status":"succeeded","stepsUsed":2,"completedSteps":1,"finalUrl":"https://example.com/","data":{"heading":"Example Domain"},"errorCode":null,"chargedSteps":0}

Errors are machine-readable: INVALID_INPUT, MISSING_API_KEY, NETWORK_BLOCKED, TASK_TIMEOUT, TASK_FAILED, OUTPUT_VALIDATION_FAILED, STEP_LIMIT, BUDGET_LIMIT and TARGET_RESTRICTED. A successfully exited Apify run can still contain success=false: inspect the final row. On platform kill, final output may be unavailable. A restart does not automatically repeat browser actions or uncertain billing/dataset writes.

Cost and billing

PPE pricing: USD 0.01 per successfully executed browser action, plus the standard Apify startup event of USD 0.00005 per GB (USD 0.0001 with the default 2 GB). AI provider usage is separate and uses your own key. Check the Pricing tab for the effective rates. Event: completed-browser-step. No completion or dataset-row event is configured. With default memory, 1/5/10 completed actions cost USD 0.0101/0.0501/0.1001 including startup, excluding provider usage. A zero-action result incurs only the startup event.

Completed error-free browser action steps qualify; done and wait do not. Prior valuable steps can remain billable if the final task later fails. Failed actions are not charged. Application-level ambiguous charging requests are recorded pending and not automatically replayed. SDK/remote retry behavior still needs verification. A live two-page test reconciled one completed action with one charged event. Broader maximum-spend, retry and abort reconciliation remains unverified. This is a beta; choose a small platform spending limit.

Security and limitations

Fresh temporary browser profile, no user browser profile or saved login. Private destinations are blocked through a public-IP-pinning forward proxy, plus browser route/allowlist checks. WebSockets/service workers/downloads and Browser Use code/file/search/extraction tools are disabled. No arbitrary API endpoints. Telemetry/cloud sync are disabled. No authenticated workflows or credential persistence in this candidate. No CAPTCHA or access-control evasion. 403/404/429 return a target limitation. Strict resource allowlisting can break sites with third-party assets. Screenshot heuristics do not constitute a general sensitive-image detector; leave screenshots off for sensitive targets. Framework/provider logs are suppressed; raw exception messages and model reasoning are not emitted. The candidate still requires broader adversarial, auth-artifact, timeout/pages/memory, upstream proxy and live-provider testing before Store release.

Local development

Python 3.12. Install requirements-dev.txt in an isolated virtual environment, then install Chromium with Playwright. Run: python -m pytest -q Browser smoke: python -m tests.browser_smoke Scripted integration: python -m tests.integration_scripted Actor entrypoint: python -m src Apify CLI deployment uses development tag. Docker base image is pinned by digest. Docker is absent locally; actual container validation is performed on Apify.

Three task-example drafts

1. Read a public page heading into JSON: examples/heading.json.
2. Extract public quote text: examples/quotes.json.
3. Compare two public pages: examples/compare.json. Live-provider inputs/outputs must be verified before publishing examples. API/n8n/Make calls use the ordinary Actor run interface after validation. This Actor is not an MCP server.

Family

Puppeteer: simple deterministic Chrome workflows. Selenium: WebDriver workflows. CDP Runner: diagnostics/emulation. Playwright: modern deterministic actions. HAR Inspector: network diagnostics. Browser Use: AI goals. MCP Server: interactive tool calls.

Support

Provide run ID, version, sanitized input and expected/actual behavior. Never include keys, passwords, cookies, auth headers or sensitive screenshots. Store title: Browser Use AI Task Runner | Natural-Language Web Automation. Keywords: browser use, AI browser agent, web automation agent, natural language browser automation, browser AI, structured web task, AI web agent. Categories proposed: AI, Automation. See the Actor page for current publication status.