For the complete documentation index, see llms.txt. This page is also available as Markdown.

Internal APIs Are All You Need

Companion docs for the Unbrowse whitepaper.

  • Authors: Lewis Tham, Nicholas Mac Gregor Garcia, Jungpil Hahn

  • arXiv: 2604.00694

  • Paper draft synced here: April 1, 2026

Important The PDF mixes research results, product behavior and forward-looking economic design. These docs separate the three: what the paper measured, what the hosted Unbrowse service does today, and what is still roadmap.

The short version

Most of the web's value sits behind interfaces built for humans. Websites already call their own internal APIs.

Unbrowse learns those requests once, in a cloud browser, and compiles them into capabilities agents call over HTTP. Later agents skip the browser.

What ships today

Unbrowse is a hosted service (https://unbrowse.ai, closed source) with an open-source client (this repo).

The service:

  • records tasks in a cloud browser (patchright) and fulfils the first request while learning it

  • compiles recorded sessions or HAR files into learned.* capabilities (harness YAML, unbrowse/v1alpha1)

  • replays them over first-party HTTP, with a render fallback for reads HTTP cannot finish

  • verifies outcomes: HTTP 200 is not success

  • keeps a public registry of scrubbed, re-verified read-only site tools, each site also an MCP server and an OpenAPI document

  • keeps logins in a password manager the model never reads, signs in on replay and reuses sessions

  • bills only verified calls: 500 a month free, then $10 per 10,000

  • routes plain-language tasks with Jev, with a deterministic fallback

The client is @unbrowse/sdk and a thin CLI over the service's REST API (/api/v1); agents use the remote MCP (/mcp).

Not built: per-route pricing, contributor payouts, validator markets, attestation.

Pages

Where to start

Citation

Last updated