Scout Browser Agent

Give an AI model a browser and a task. It opens pages, reads them, and reports what it found, using your own model provider or AI gateway key, so the run spends your account for the model and Scout's pricing covers only the browser.

How it works

1. You give it a task, a key, and optionally the domains it may visit.
2. A model — yours — is handed a set of browser tools and works through the task one step at a time.
3. When it has an answer, or runs out of steps, budget or time, it writes one result to the dataset.

Hitting a limit is never a failure: the model is asked to say what it found so far and what it could not find, and that is the result. Only a run with nothing at all to report fails, and a run that fails is not charged for a result.

Input

Field	Required	What it is
prompt	yes	The task, in plain words.
apiKey	yes	The key for the account model starts with. Stored encrypted, used only to call that account, never shown to the agent or logged.
model	no	Whose key the run spends, then the model as they name it: anthropic/claude-sonnet-5-5, openrouter/anthropic/claude-sonnet-5.5. deepseek/deepseek-flash when not given.
baseUrl	no	For a custom/… model only: the https address of an OpenAI-compatible endpoint.
allowedDomains	no	Domains the agent may navigate to, with their subdomains. Empty means any public page.
maxSteps	no	The most model turns before it must answer (1–50, default 25). A turn can run several tools. Turns are not charged; tool calls are.
tools	no	Browser tools by name, picked from a list in the form. Empty means the default set below.

The browser's own settings (timeouts, snapshot limits, proxy) are further down the input form and have sensible defaults.

Providers and gateways

model starts with whose key the run spends. A provider is the vendor that makes the model, called directly with that vendor's key: anthropic/claude-sonnet-5-5. A gateway is a router in front of many vendors, called with the gateway's key, and what follows it is the name the gateway uses: openrouter/anthropic/claude-sonnet-5.5. One value carries the whole route, so a model cannot end up on the wrong account.

Because the key is yours, every provider Scout can reach is available:

Provider	model starts with	Spoken to
Alibaba, Arcee, ByteDance, DeepSeek, Interfaze, Mistral, Moonshot, OpenAI, Sakana, StepFun, Tencent, xAI, Xiaomi, Z.ai	alibaba/, arcee/, bytedance/, deepseek/, interfaze/, mistral/, moonshotai/, openai/, sakana/, stepfun/, tencent/, xai/, xiaomi/, zai/	Chat completions, at the address Scout keeps for each
Google (Gemini)	google/	Google's OpenAI-compatible endpoint
Cohere	cohere/	Cohere's OpenAI-compatible endpoint
Anthropic	anthropic/	Anthropic's own Messages API
Anything else	custom/, with baseUrl	Chat completions, at a public https address you give

Gateways: openrouter/, vercel/, kilo/, trustedrouter/.

Models change faster than this page can. Nothing here pins a model name except the default, which DeepSeek's own documentation tells callers to use (deepseek-flash). The input form suggests current names, each titled with the key it needs, but it is a list of suggestions and not a limit: any name your key can use works. The list comes from models.dev: only names, never addresses. To refresh it, run pnpm --filter @scouting/runtime-browser models:refresh and commit the form. Before a run starts, a key for a named provider is checked with that provider, so a mistyped one fails at once and costs nothing (a gateway or custom address finds out at the first request); and if the provider refuses a model name, the failure lists the models your key can use that are nearest to it.

{
  "prompt": "What does example.com say its purpose is?",
  "apiKey": "…",
  "allowedDomains": ["example.com"]
}

{
  "prompt": "Summarise the pricing page.",
  "model": "anthropic/claude-haiku-5-5",
  "apiKey": "…"
}

{
  "prompt": "Find the pricing page and summarise the plans.",
  "model": "openrouter/deepseek/deepseek-v4.1-flash",
  "apiKey": "…"
}

Output

One dataset item per run, also stored as OUTPUT:

{
  "answer": "example.com describes itself as…",
  "prompt": "What does example.com say its purpose is?",
  "model": "deepseek-flash",
  "route": { "provider": "deepseek" },
  "steps": 4,
  "toolCalls": 3,
  "stoppedBecause": "answered",
  "usage": { "promptTokens": 9120, "completionTokens": 310 },
  "screenshots": ["SCREENSHOT-1"]
}

stoppedBecause is answered, max-steps, budget, time, or error (the provider failed part-way, was cut off at its token limit, ran out of context, or declined; error then says which, and answer is whatever the model had said). usage is what your provider account was charged for, so you can see it without opening that account; cachedPromptTokens is the part of the input the provider served from its own cache, which is billed at a fraction of the rest.

What else a run leaves

Record	Where	What it is
TRACE	Key-value store	Every step: what the model said, each tool it called with what it was given, and what came back. Kept for failed runs too.
SCREENSHOT-1, -2, …	Key-value store	The pictures the browser took, up to 25 a run. The model cannot see them; they are for you, and the result lists their keys.

The trace shows what the model saw, which can include page content. It never contains your API key.

Limits and safety

• Allowed domains. Navigation outside the list is refused by the browser itself, including redirects that leave it. What an allowed page loads into itself — scripts, images, frames — is not restricted, or no page would render. Private and internal addresses are always refused.
• Pages are untrusted. The agent is told that page text is information, not instructions. It is given only the tools below, and your key is never in its reach.
• Your key. It goes to the one endpoint you named and nowhere else. A custom base URL must be a public https address.
• Budget. The run stops short of the maximum charge you set on it, keeping back enough to charge for the answer.
• Time. The run stops working about three minutes before its timeout (a quarter of the way before, on a short one) so that it can still answer. Give a long task a longer timeout.

Default tools

browser-open, browser-navigate, browser-snapshot, browser-find, browser-element, browser-interact, browser-extract, browser-search, browser-wait, browser-tabs, browser-close.

The rest of the browser's tools can be selected in the form or named in tools. Those left out of the default reach further than reading does — rewriting request headers, running script in a page, cookies and storage, files — so give them deliberately. The form marks the ones to be most careful with as Sensitive.

Pricing

Pay per event, with Apify platform usage included:

Event	Price	When
tool-call	$0.01	Each time the agent runs a browser tool.
result	$0.05	Once, for the answer the run ends with.

Model tokens are not part of this: they are billed by your own provider or gateway.

All browser tools (70)

Every tool the browser offers. Select the ones you want in the form, or name them in tools; the agent finds sessionRef and tabRef (which identify a browser session and a tab) on its own.

Navigation & Page Control

Tool	Description
browser-launch	Launch a new local Chromium with extensive configuration (flags, geolocation, locale, color scheme, network, devices, emulation)
browser-open	Launch Chromium with a persistent user-data directory for saved sessions, cookies, and profiles
browser-connect	Connect to an existing Chrome/Chromium instance by domain and port
browser-remote	Connect to a remote Playwright browser server via WebSocket endpoint
browser-close	Close the current browser session or all tracked sessions
browser-disconnect	Release a tracked Playwright browser session
browser-navigate	Page navigation — goto, reload, back, forward (with waitUntil options)
browser-wait	Wait for conditions — load state, delay, selector, text, URL, JS function, navigation
browser-search	Run a search query through Google, DuckDuckGo, Bing, Brave, Yandex, Baidu, and more
browser-find	Text-search within a browser tab, returning match count and position
browser-resize	Resize the browser window to specified width and height

Tabs & Frames

Tool	Description
browser-tabs	Manage tabs — list, create, switch, close, activate
browser-frames	Manage iframes — list frames, switch between them

Interacting with the Page

Tool	Description
browser-interact	Full interaction suite — check, clear, click, dblclick, focus, hover, fill, press, scroll, selectOption, setChecked, tap, type, uncheck
browser-pinch	Perform a pinch gesture at a page coordinate to zoom in/out
browser-attach	Attach files by targeting a file input element with local file paths or file refs
browser-inject	Inject inline JavaScript or CSS into a page (inject/status/remove lifecycle)
browser-evaluate	Evaluate a JavaScript expression in a tab context, optionally targeting a specific frame

Reading the Page

Tool	Description
browser-snapshot	Snapshot the page's accessibility tree with intelligent filtering — deduplication, disambiguation, role filtering, depth flattening
browser-screenshot	Capture a screenshot (JPEG or PNG, with optional ref annotation overlay)
browser-pdf	Generate a PDF from the current page with layout and formatting options
browser-extract	Extract values from the DOM — frame properties (article, HTML, markdown, text, title, URL) or element state (box, checked, enabled, focused, HTML, markdown, text, value, visible, attribute)
browser-source	Capture document HTML, stylesheet sources, script sources, or image asset metadata
browser-styles	Retrieve computed CSS styles for a targeted element
browser-image	Read a snapshot-issued image ref and return it as PNG or JPEG
browser-clipboard	Read or write clipboard text in the browser session
browser-element	Inspect one element by @ref or selector: position, size, visibility, content, attributes, styles, accessibility, state or value
browser-layout	Overlay a layout-debugging view on one element and return its geometry, optionally with measurement guides

Content Extraction & Export

Tool	Description
browser-crawl	Multi-URL crawling with site-wide configuration, resource blocking, browser diversity, and result collection
browser-pipe	Evaluate a browser expression and write the result to a file in YAML, JSON, text, or base64 format
browser-download	Track and manage file downloads — get, list, clear

Console, Network, Storage

Tool	Description
browser-console	Inspect browser console logs — clear, get, list with type and pattern filtering
browser-exceptions	Inspect uncaught page exceptions — clear, get, list with pattern filtering
browser-network	Track and inspect network requests — list, get with URL pattern filtering
browser-route	Intercept network requests — abort, continue (with header override), fulfill, fallback
browser-unroute	Remove one or all previously registered network interception routes
browser-har	Control HAR recording — start, stop
browser-throttle	Apply network throttling — latency, throughput, offline mode
browser-security	Control TLS/certificate validation error handling
browser-cookies	Manage cookies — get, set, delete, clear
browser-storage	Access localStorage or sessionStorage — get, set, remove, clear
browser-request	Issue a controlled request from the session network layer and return a bounded raw response; the network guard applies to every redirect hop

Browser State & Metadata

Tool	Description
browser-history	Inspect or clear recorded navigation history
browser-scene	Save and restore page state (URL, title, scroll position) — capture, restore, list, delete
browser-emulate	Emulate device, locale, timezone, and User-Agent Client Hints
browser-metrics	Retrieve performance and runtime metrics
browser-memory	Inspect runtime memory state
browser-heap	Capture heap information, optionally forcing garbage collection first
browser-cpu	Adjust CPU throttling to simulate slower device performance
browser-trace	Control Chrome performance trace recording with configurable categories
browser-file	Inspect/clear tracked file references used for uploads
browser-probe	Declare a bounded experiment against the current page, so what is done to it is attributable to one named experiment

Media & Accessibility

Tool	Description
browser-media	Control and inspect media elements — info, mute, unmute, play, pause, seek
browser-transcribe	Extract raw audio bytes from video or audio elements as base64
browser-still	Capture a still image from a media element at a specific timestamp
browser-screencast	Capture video frames or MP4 screencast with configurable interval, format, and resolution
browser-vision	Apply vision-deficiency filters for accessibility testing — blurred vision, reduced contrast, achromatopsia, deuteranopia, protanopia, tritanopia
browser-highlight	Visual highlighting with measurement overlay — element bounds, size, viewport distances, border-radius

Dialogs & Notifications

Tool	Description
browser-dialog	Handle browser dialogs — accept, dismiss, dismissAll, get
browser-notifications	Retrieve browser notifications, optionally clearing them

WebSocket & Runtime

Tool	Description
browser-websocket	Inspect/clear tracked WebSocket connections with pattern filtering
browser-animation	Control animation playback — pause, resume, set playback rate

Canvas

Tool	Description
canvas-diff	Compare two canvas or screenshot images pixel by pixel
canvas-inspect	Inspect a canvas's dimensions, bounds, contexts and pixel export
canvas-interact	Point-based pointer and keyboard input, or low-level events, inside a canvas

Extensions

Tool	Description
extension-launch	Launch Chromium with an unpacked extension loaded
extension-open	Open an extension page (popup, side panel, options) in a new tab
extension-background	Evaluate JavaScript in the extension's background service worker
extension-reload	Reload the extension and wait for its service worker to come back
extension-storage	Read or write the extension's chrome.storage areas