Vision

Vision is how your agent reads and uses the web. It turns pages into clean Markdown and other formats, discovers and crawls sites, and gives your agent a browser of its own to click through, all from browsers and contexts that are already running when the request arrives.

What to call

How a scrape is served

A scrape has three ways to get a page, and engine: "auto", the default, takes the cheapest one that works:

1. Cache. A copy no older than max_age_ms (two days by default) answers at once.
2. Fetch. A plain HTTP request, with the page converted to Markdown in Rust inside the B3OS daemon.
3. Browser. A warm Chrome page from a pool of pre-created contexts, used only when the page needs JavaScript to show its content.

engine: "fetch" never starts a browser. engine: "browser" always renders, and browser actions, screenshots and profiles imply it. The response's tier says which one served the page; each bills the same single vision.page.

The cache

A page enters the shared cache only when the request had no custom headers or cookies, no profile, session or browser actions, the default location, and the site set no cookies. Anything else is cached for your organization only, or not cached at all. cache.state in the response is hit, miss (fetched and stored) or bypass (fetched and not shared), and cache.age_s is the copy's age. Cached pages are kept for two days. Set max_age_ms: 0 to always fetch, or store_in_cache: false to keep a result out.

Browsers

Two kinds of browser serve vision, and they never mix:

• The scrape pool serves scrapes, batches and crawls that need JavaScript. Its browsers are driven only by B3OS; nobody connects to them over CDP.
• Session browsers are whole browser processes, one per session with its own profile directory, claimed from a warm pool. Your agent drives one with batched actions or any CDP client, through a proxy that exposes only that session's pages. During the beta Chrome runs without its renderer sandbox; the beta contract says what that means.

Both are Chrome headless shell by default. All their traffic goes through the egress guard.

Rules vision follows

• robots.txt and Content Signals are honored by default. A URL they disallow fails with robots_disallowed; the beta has no override.
• Politeness. Crawls space their requests to one host.
• Egress. Every fetch and navigation goes through the egress guard, so a URL that resolves to a private, loopback or metadata address fails with egress_denied.
• Ads and trackers are blocked by default in scrapes (block_ads), and images, media and fonts are not downloaded unless you ask for a screenshot.

Scopes

Guides

Limits and FAQ lists the caps and the common questions. Firecrawl users can keep their SDK: see Firecrawl compatibility.