Organization repository on GitHub
Plug any CDP browser into Jev — cloud, local or self-hosted, including browsers that never draw a page.
- Forks
- 0
- Last pushed
- Updated Oct 2026
Repo stats from the GitHub API, cached Oct 2026.
CommunityJev Demos & Experiments
Jev Browser Bridge
Agent bridge that connects any CDP-speaking browser to Jev by reading the DOM instead of the rendered layout, so headless or non-drawing browsers can still drive CLICK, TYPE_TEXT, SELECT, WAIT, DONE, and BLOCKED steps.
- Stars
- 10
- Language
- Python
- License
- APACHE-2.0
- Published by
- Community
- Author
- lexmount
- Last verified
- Oct 2026
About this project
Cloud, local or self-hosted. Chromium or not. Even browsers that never draw a page. If it speaks CDP, it runs Jev. One task, one Jev model, four browsers, each read two ways — recorded runs on a shared clock. Pictures are each browser's own screenshot; Lightpanda answers with a placeholder and Obscura has no screenshot at all. The position-based reader finishes on the browsers that draw pages; Jev Browser Bridge finishes on all four . Open the player to pause, scrub and read each run's raw trace.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Each step builds an element table and asks Jev for operation plus target in one request.
- Clicks dispatch on the element itself, so there is no coordinate to go stale.
- Long pages are split into rows and retrieved against the goal up to a character budget.
- Same-named controls are labelled by location; script-bound spans surface via pointer styling.
Watch out
A small text model is only needed for goals that type into a field, and MiniWoB++ tasks needing pointer position, drag, drawing, colour, or password are excluded. Copy reviewed against the repository README on 2026-10-04, not from running it.
Questions about this entry
Short answers drawn from the linked project or page.
A position reader depends on the layout engine and viewport, so it shrinks on lazy, missing, or fake rendering. This bridge reads DOM attributes, stylesheet pointer rules, and retrieved page rows, which is why it completes tasks on browsers that draw on demand or never draw.
Every observation becomes an element table and Jev picks the operation and its target in the same request, with only observed elements offered. A small text model writes a string only when the operation is TYPE_TEXT.
Same-named controls are labelled by where they live, links sharing a name and destination are offered once, and script-bound spans are found through pointer styling or parsed style text. A field with no label is named by adjacent text.
Any CDP endpoint works. Lightpanda and Obscura never draw pages and the bridge still finishes on them. Pass the address to connect(), or --cdp as an http:// or ws:// URL, and use lexmount_session() for Moli. JEV_API_KEY goes in .env; TEXT_MODEL_API_KEY is only needed for goals that type into a field, and pytest runs the retrieval and retry checks with no browser and no key.
MiniWoB++: 86 tasks x 2 seeds = 172 seeded episodes per browser, with the bridge scoring 106/172 on Chrome, 112/172 on Browserbase and 110/172 on Lightpanda, versus 54/172 for the position-based reader on Chrome.
One request per step chooses the operation (CLICK, TYPE_TEXT, SELECT, WAIT, DONE, BLOCKED) and the target for each; only observed elements are offered, and a small text model writes a string only for TYPE_TEXT.
A long article is 50,000-150,000 characters; the page is split into rows and the rows bearing on the goal are kept in page order up to JEV_EVIDENCE_CHARS, default 20,000, since a 6,000-character prefix missed the answer every time.
More Jev resources like this
62GitHub stars
Simulation study that drives one xArm7 apple-to-plate task with Jev, a large reasoning model, and a small model through identical incremental controls.
Jev Demos & Experiments#community#jevbest#catalog-import
114GitHub stars
Browser agent where Jev picks each action from indexed page candidates while an LLM plans and reads, returning cited answers or schema-checked data with credential controls.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
agent-labs-dev/fastbrowsePython
agent-labs-dev/fastbrowsePython4GitHub stars
Browser-use layer for CLI coding agents: Jev picks the operation and target element in one round trip per step, exposed over MCP with a CLI, a Python API and an HTTP server.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
brnyxx/jev-raPython
brnyxx/jev-raPython2GitHub stars
A Python decision layer for browser and desktop automation: your code observes the surface, builds the candidate list and verifies the result, and the model only picks one id per step.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
aidil2105/jev-browser-pilotPython
aidil2105/jev-browser-pilotPython0GitHub stars
A local Python browser automation UI that turns a URL and a plain-language goal into bounded actions, with Jev picking the next operation from observed candidates and a separate text model used only to fill fields.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
KesavanKing/jev-browserPython
KesavanKing/jev-browserPython1GitHub stars
A Streamlit Tetris where the engine enumerates every legal placement, Jev picks one typed placement ID, and Python verifies rotation, column and exact occupied cells before the piece is allowed to lock.
Jev Demos & Experiments#community#jevbest#catalog-import
From the Jev guides
Original write-ups on what this entry is for.
A game tick is a state plus a legal action list. How Clash Royale, Minecraft, and the earlier Mario, Pokémon, StarCraft, and drone loops ask Jev one step at a time.
Jev Demos & Experiments#gaming#agents#control
Read article
From the Jev community
Posts from builders shipping with Jev right now.