Your agent already reads the web.
Badly.

Search everything companies publish, by meaning and by words at once. Use this first for any question about what a company says, how an industry writes about a topic, or where a claim comes from. Returns the matching SECTION of each page, not the whole page, so you can quote it directly. Filters apply inside the search rather than after it, so narrowing does not silently reduce the result count.

q · domain · collection · type · locale · technicality_min · promotional_max · mode · limit

Browse pages by their PROPERTIES rather than by a query — 'technical articles with code from developer-tool companies', 'pricing pages changed this year', 'the most-cited posts on this domain'. Use this when the question is about a category of writing rather than a subject. Every filter is a value scored when the page was indexed; none of them is a text match. Two quality gates apply unless you turn them off: pages under 150 words and domains with a spam score of 50+ are excluded. The response declares them in `qualityGates` with a count of what they removed, so a short result is explained rather than ambiguous.

domain · collection · type · format · audience · topics · technicality_min · depth_min · originality_min · promotional_max · has_code · include_thin · published_after · sort · limit

Start here when the question is about a CATEGORY of company rather than a named one — "the docs of the major payment APIs", "what shipped across AI providers this week". Each shelf is a set of companies somebody chose, with its own refresh cadence: documentation weekly, changelogs daily, open source monthly. Pass a slug as `collection` to search_content or list_content to scope a query to one. Read `unindexed` before quoting a shelf as complete: it counts members we hold nothing for yet.

slug · limit

The full cleaned markdown of one page, plus what we scored about it, its images with their alt text, every crawl we have of it, and which of the company's own pages cite it and under what wording. Use this after a search when you need the whole argument rather than a section, or when you want to know whether a page has changed.

id · markdown

What a company publishes, in what languages, how much of it we have, and what moved recently. Includes their own sitemaps, pages that were redesigned without the words changing, and which other companies they link to. Read `coverage.headline` before quoting any total: if we only have part of a site it says so, and says which bucket the rest is in.

domain

Start crawling a domain. Returns a job id immediately — a crawl takes minutes to hours, so poll job_status rather than waiting. Check get_domain first; if it returns a result the site is already indexed and you want that instead.

domain · budget

Progress on a crawl: indexed, pending, discovered, elapsed, pages per minute and an eta. `coverage.buckets` accounts for every URL found — `unaccounted` is always 0, and pages left unfetched sit in `budget` or `stoppedEarly` with the reason attached. A finished crawl is not the same as a complete one; read the buckets.

jobId

Do I need an account or a key?

No. Every read returns exactly what the public website already shows, so a key would buy nothing and would cut the server out of anonymous discovery — which is most of the reason to run one.

What does it cost?

Nothing. index_domain is the only tool that spends our money and it is bounded by a per-domain page budget.

Is this the same as giving my agent a browser?

A browser fetches one URL and hands back the navigation, the cookie banner and the footer with it. This is 405 pages already converted to clean markdown, split on headings, embedded, and searchable across 10 companies at once. The difference shows up the first time the question is broader than one page.

What is in the corpus?

405 pages from 10 companies in 89 language(s) — blogs, docs, pricing, changelogs, landing pages. A translated page is a separate document, not a variant, so a Spanish docs page is its own row.

Can I add a company that is not there?

Yes, and that is what index_domain is for. It returns a job id immediately because a crawl is minutes to hours; poll job_status rather than waiting on the call.

How current is it?

Every domain is on a refresh schedule and every fetch is diffed against the page's own past. get_domain returns what moved recently, including pages that were redesigned without a word changing.

What happens if you only have half a site?

It says so. coverage.headline is generated in SQL from a view with no write path, so the site, the API and this server cannot word it three different ways — and the pages we do not have sit in a named bucket with a reason attached.