fastbrowse

**A browser agent that picks instead of generating.**

Jev chooses each action, an LLM plans and reads, and code owns verification, safety and secrets.

[](https://pypi.org/project/fastbrowse/)




</div>

> **Warning**  
> This project is highly experimental and not recommended for production use yet.

---

## Why

Most browser agents generate each action from a screenshot. fastbrowse indexes the page into
candidates and has [Jev](https://typesafe.ai), a choice model, **pick one**, so it cannot click
something that was never on the page. Every claim in an answer cites a verbatim quote from the page.



### Against Browser Use

The same 14 answer tasks (lookups, sign-ins, checkout, Google Flights), three passes each, on the same kind of
cloud browser, run the same day with **no cost cap on either side**, so a run stops when the agent is done
rather than when a budget runs out. Neither arm hit a limit.

| | passed | cost per task | median time |
|:--|:--|:--|:--|
| **fastbrowse** | 41/42 | **$0.0057** (median), $0.0084 mean | **21.0s** |
| Browser Use (hosted) | 42/42 | $0.41 (median), $0.53 mean | 24.8s |
| | | **71x cheaper** | 1.2x faster |

The whole suite cost $0.35 here and $22.28 there.

**Reliability is a tie, and the earlier claim was wrong.** An earlier version of this table read "2.9x the
passes", from a run where both arms were given a $0.25 cap. Browser Use exceeded that cap on 28 of 42 runs and
each was scored a failure. Uncapped it passes every one: 36 of its 42 runs cost more than $0.25, so the cap was
measuring our budget, not its ability. We have no reliability advantage to claim on this suite, and the one
failure in the table is ours (`saucedemo-locked-out`, one pass of three).

**Where the cost difference comes from.** It grows with how much a task do