jev-browser (MahmoudAdelbghany)
Jev-powered browser MCP server that makes each step in one parallel call and escalates to the driving LLM only when unsure.
Overview
This jev-browser is a browser MCP server for LLM agents in which Jev picks the next browser action in about 300 ms with no LLM tokens, and the driving agent only steps in on escalations. Each step takes a compact snapshot of candidates ranked locally by goal-token overlap, then makes one Jev call with parallel questions covering action, element, key, field-to-text binding, whether the previous step succeeded, and whether the goal is achieved. Actions are performed as trusted CDP input, and a deterministic state diff of URL, text and inputs verifies progress, with Jev's prev_ok only a tie-breaker.
Because Jev cannot generate strings, typing uses a joint ref|text choice to bind a value to a field in one pass. Invalid combinations get one corrective re-ask, and low confidence, stuck states or loops return an escalation with page state, ranked candidates and reasons. Against Playwright MCP with the same driver model on a 12-task suite, the README reports 1.5x faster and 1.6x cheaper at the same accuracy, and the autonomous Jev loop alone at about 1.8 s and $0.0005 per task with 97% success.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which action and element should be used next? | — |
| Choice | Which field gets which supplied text (ref|text)? | — |
| Yes / no | Did the previous step succeed / is the goal achieved? | yes · no |
How it works
- Snapshot page into a ranked candidate list
- One Jev call answers action, target, binding, prev_ok and goal
- Act via CDP and diff the state
- Continue if confident and verified; otherwise escalate to the driving LLM
Strengths
- One call per step folds in verification and done-detection
- Deterministic verification overrides weak model signals
- Reported 1.5x faster and 1.6x cheaper than Playwright MCP
Caveats
- Small 12-task suite
- Requires a Jev early-access key
Getting started
`git clone https://github.com/MahmoudAdelbghany/jev-browser.git && cd jev-browser && npm install && npx playwright install chromium`, then set the key in .env.At a glance
- Type
- Benchmark
- Status
- beta
- Language
- TypeScript
- Runtime
- MCP server (Playwright, Node 20+)
- Has code
- yes
Use cases and integrations
How curated lists describe it
Jev-powered browser MCP for LLM agents — ~300ms decisions, no LLM tokens in the loop. Benchmark vs Playwright MCP included.
hellogumbo/awesome-jev
Frequently asked questions
How does it type text if Jev cannot generate strings?
It offers joint field-and-text options and Jev picks the right pairing.
When does the driving LLM get involved?
Only on low confidence, stuck states or loops.
How does it compare to Playwright MCP?
1.5x faster and 1.6x cheaper at the same accuracy on 12 tasks, per the README.
Related projects
Reproducible bake-off of Playwright CLI alone, CLI plus Jev, and Playwright MCP on fuzzy browser goals, with honest results.
MCP serverresearchPythonMCP server and libraries that drive Chrome over CDP and hand a decision model like Jev a bounded menu of page actions.
ProjectproductionTypeScript- Jev Browser★ 335
Headless browser automation via MCP server, CLI or library: Jev picks one action per step and scores goal-met and stuck probabilities.
MCP serverbetaTypeScript Claude Desktop MCP extension that drives Brave or Chrome with Jev choosing on-page actions and a guard freezing irreversible clicks.
MCP serverbetaTypeScriptBrowser-agent benchmark with five timed events from one prompt, plus a ~200-line DevTools agent where Jev picks each action.
ProjectbetaJavaScriptChrome MCP server with 50 tools that drives your everyday browser without losing logins, where Jev or a local decider picks each step.
MCP serverbeta