Headless browser for AI agents

One call reads the page

Send the URLs and get the text. There is no session to start, keep alive or close, and no browser to pay for while the model thinks.

curl https://api.specstream.ai/v1/extract \
  -H "Authorization: Bearer $SPECSTREAM_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"urls": ["https://example.com/pricing", "https://example.com/changelog"], "format": "markdown"}'

From an agent, connect the MCP server at https://api.specstream.ai/mcp and call read_pages with {"urls": [...]}. The setup for each client is at MCP.

A real browser when the page needs one, at the page price

Each URL is read plainly first. A page that refuses that read is tried again as a browser would ask for it, then in a real browser, then from the site's own feed, then from its newest public archive capture. Each result says where it was read from in read_from: live, site or archive, with captured_at for an archive copy. Every page read costs $0.0008, whichever way it was read.

You pay for the text you get

A page that gives no text is not charged. Neither is a 404 page, a plain 403 or a bot check: the status says not_found, blocked or failed, and a bot check's page is never returned as text. At $0.80 per 1,000 pages, reading costs less than Browserbase's Fetch with Extract at $4 per 1,000, Steel's /scrape at $5 per 1,000 and Hyperbrowser's Fetch with JSON output at $6 per 1,000.

Text a model can use

text holds the page's main text, without its menus, footers and ads, as Markdown with headings, lists and links kept, or as plain text. Each result also carries title, published and modified when the page states them, and emails with every address the page carries, also those only its markup shows. PDF, Word and Excel files are read the same way. The Extract API scores 0.908 F1 on WCXB, a test of how well an extractor keeps a page's main text and drops the rest (benchmarks).

Search and structured rows on the same key

The agent finds the pages with the Search API, at $5 per 1,000 searches of up to 10 results or $1 per 1,000 on the fast tier, scored 97.9 on SimpleQA Verified and 64.0 item F1 on WideSearch. When it needs many records with the same facts, such as every company that meets a condition, Deep Search returns them as structured rows at 5¢, 15¢ or $1 a row, with a source URL and read time on every value. One key and one balance pay for all three.

When a step has to click or sign in

A form, a sign in or a checkout stays in the browser your agent already runs, such as Playwright or Browser Use, on your own machine. SpecStream then does the searching and the reading, so the browser opens only the pages it has to act on. The Browser Use guide adds SpecStream search to a browser agent in one function.

Browserless

Hosted browsers billed in units: a unit is up to 30 seconds of browser time per connection. Prototyping is $25 a month billed annually for 20,000 units, then $0.0020 a unit. On SpecStream a page read costs $0.0008 however long the site takes, up to 30 seconds.

Playwright

A free, open-source framework under Apache 2.0 that drives Chromium, Firefox and WebKit with one API, run on your own machines. You keep the browsers, the addresses and the retries. SpecStream reads the pages your agent only has to read, with the browser on its side.

Hyperbrowser Fetch

Hyperbrowser's Fetch API reads a page for $0.001, or $0.006 with JSON output. Its free account holds 5,000 credits. SpecStream reads a page for $0.0008 with the real browser included, and returns its main text with dates and emails as named fields.