jev-browse
Helpers for browser-harness that let Claude Code or Codex hand off a whole website sub-task in one call, with Jev deciding each step.
Overview
jev-browse adds helpers to browser-use's browser-harness so a coding agent such as Claude Code or Codex can hand off an entire website sub-task, like searching for Lisbon, setting two filters and opening a listing, in a single fast_run call. Each step is decided by Jev in a few hundred milliseconds, and the agent gets back the outcome or a typed reason it stopped. It hands control back only when necessary: a value it does not know, a risky click, a frame or upload it cannot handle, or anything sensitive.
The motivation is that when an agent drives a browser itself, every click costs a model turn that re-reads the whole context, so a five-field form becomes fifteen turns. In the author's runs this made form and navigation tasks 1.6-3.0 times faster and 2.1-5.3 times cheaper with the same pass rate; a Google Flights benchmark (median of three, 24 September 2026) shows 72.8 s and $2.34 versus 24.7 s and $0.44. It is published on PyPI with CI, coverage and OpenSSF Scorecard badges, and MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which page element and action advances the sub-task?(inferred) | — |
| Yes / no | Must control be handed back (unknown value, risky click, sensitive step)?(inferred) | yes · no |
How it works
- Agent calls fast_run with a URL and goal
- Jev decides each browser step
- Stop and return a typed reason if input or caution is needed
- Return the outcome to the agent
Strengths
- Reported 1.6-3.0x faster and 2.1-5.3x cheaper than agent-driven browsing
- Same pass rate in the author's runs
- Typed stop reasons
Caveats
- Benchmarks run on the author's machine with small N
- Depends on browser-harness
Getting started
`pip install jev-browse` and call `fast_run(url, goal, run_id=...)` inside a browser-harness script.At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- browser-harness helper library
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Fast, cheap browser sub-tasks for Claude and other agents: TypeSafe Jev decisions on top of browser-harness
logicrw/awesome-jev-projects
Frequently asked questions
What does fast_run do?
Completes a whole website sub-task in one call, with Jev deciding each step.
When does it hand back to the agent?
For unknown values, risky clicks, unsupported frames or uploads, and sensitive steps.
How much faster is it?
1.6-3.0x faster in the author's runs.
Related projects
MCP wingman that hands an agent's browser clicks and typing to Jev on a real logged-in Chromium, alongside Playwright MCP.
MCP serverbetaTypeScriptmacOS computer-use loop that OCRs the screen, has Jev pick the next action from a short list, and clicks, at about $0.0002 a step.
ProjectbetaPythonLets coding agents run whole goals in your signed-in Chrome on macOS in background tabs, asking Jev or Claude Haiku only for real choices.
ProjectbetaLocal compaction proxy for Claude Code and Codex that compacts more often with local rules and keeps full history searchable.
ProxybetaGoSkill that makes Claude Code and Codex split AI features into Jev, LLM or plain-code steps before writing code.
ProjectbetaAgent-plugin browser tool where Jev picks one operation and one row from a numbered page table; a small LLM only writes typed text.
ProjectbetaTypeScript