Teach a website to your agent once, then skip the browser forever

What it does

You point it at a page with a placeholder and two example inputs. A headless browser loads the page once per example, records the requests the site’s own frontend makes, and figures out which parts of the requests changed with your inputs — those become parameters, the rest becomes a stored template. The saved operation is then callable over MCP, a CLI, or TypeScript, and returns JSON. No LLM is involved in building the template.

The interesting bit

The self-healing. When a site rotates an ID or moves a field, the tool re-learns the request and only saves the new version after a successful replay — otherwise it returns an error rather than someone else’s data. Sessions are handled by storing references to cookies and tokens, never values, so refreshed sessions keep working without re-learning. Login usually just imports cookies from a browser profile you’re already signed into.

Key highlights

• Calls fall back through tiers: plain HTTP first, then fetch from inside a browser page, then running the page itself — tier: 1 means no browser at all.
• Ships with 22 read operations across 9 sites, including Google Flights, Hacker News, Goodreads (no public API since 2020), and LinkedIn with your own login.
• In its benchmarks, an agent using the operations answered 1.5–2× faster than one driving Chrome, reading about half the tokens; a Goodreads run cost $1.19 to build and produced four operations.
• A plain script with no model built the same five-book Goodreads table in 16 s.
• Errors come back structured: ok: false with a class, a reason, and a suggested next step.

Caveats

• Local tool, not a hosted API — it uses your machine, your network, and your sessions, and only knows the operations it was taught.
• No CAPTCHA solving; if a site challenges real Chrome, you clear it by hand. Writes are tested only against a local fixture and need an explicit flag.
• Not on npm yet, and the README warns that installing directly from GitHub fails under npm 11.

Verdict

Worth a look if your agents burn tokens and patience driving browsers for data that’s really just HTTP underneath. Skip it if you need hosted infrastructure, write operations at scale, or sites that don’t survive without a CAPTCHA.