dejevu
Browser agent that acts on one look at the page with one open-model call per action; Jev is an optional chooser backend.
Overview
dejevu is a fast browser agent built as a response to Jev-based browser demos. Each step takes one look at the page, makes one call to an open model and performs one typed action. The default backend is any OpenAI-compatible open model, such as llama-3.3-70b on Groq via OpenRouter, and a --backend typesafe flag makes Jev the chooser instead.
The README compares it with Browser Use's jev-ultrafast on a Google Flights search: dejevu completed the verified task in about 5.6 seconds with 10 model calls versus a published 7.1 seconds and 17 calls for Jev plus a second model, with 5.6 times fewer tokens. Every run is verified by code on the final page rather than by the model saying it is done. It is relevant to the Jev directory mainly as a comparison point showing that a single generative call per action can compete on speed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which typed action on which page element should be taken next?(inferred) | — |
How it works
- Observe the page once
- One model call (open model or Jev) picks a typed action
- Execute the action in Chrome
- Code verifies the final page
Strengths
- Few calls and tokens per task as reported
- Works with any OpenAI-compatible endpoint
- Code-verified success
Caveats
- Default path does not use Jev
- Comparisons are on a small number of tasks
Getting started
git clone the repo, uv sync, add OPENROUTER_API_KEY to .env, uv run dejevu --doctor, then uv run dejevu --task wikipedia.At a glance
- Type
- App
- Status
- prototype
- Language
- Python
- Runtime
- CLI with Chrome
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
Jev? Déjà vu. Browser agents that run on instinct, no Jev needed. One look at the page, one call to any open model, one action. Faster than the Jev demo on Google Flights.
logicrw/awesome-jev-projectsBrowser agents that act on one look at the page; the default backend is any open model, and `--backend typesafe` makes Jev the chooser instead.
MrJev/awesome-jev
Frequently asked questions
Does dejevu use Jev?
Optionally, with --backend typesafe; by default it uses an open model.
How does it compare with jev-ultrafast?
The README reports 5.6 s and 10 calls versus 7.1 s and 17 calls on a Google Flights search.
How is success checked?
By code on the final page, not by the model.
Related projects
Browser-agent benchmark with five timed events from one prompt, plus a ~200-line DevTools agent where Jev picks each action.
ProjectbetaJavaScriptego-decision-layer: measured, pluggable decision layer for ego lite where Jev picks operation and target in one request.
ProjectbetaTypeScriptPlugs any CDP browser, including engines that never draw a page, into a Jev loop that picks operations and targets from the DOM.
LibrarybetaPythonBrowser automation where Jev picks every step from a sightmap's named actions and reports whether the goal is met.
ProjectbetaGoOpen Apache-2.0 browser-agent decision model (Qwen3.5-35B-A3B fine-tune) that doubled Jev 1.13's task success in the same agent.
ServiceresearchPythonBrowser driver where Jev makes every decision and agent-browser performs every action, benchmarked against Claude Code on 10 tasks.
BenchmarkbetaTypeScript