User repository on GitHub
Browser use for coding agents, 3-5x faster than browser-use. MCP server + CLI; TypeSafe Jev decides every step in ~300 ms.
- Forks
- 1
- Last pushed
- Updated Sep 2026
Repo stats from the GitHub API, cached Oct 2026.Homepage
CommunityJev Demos & Experiments
Jev Ra
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.
- Stars
- 4
- Language
- Python
- License
- MIT
- Published by
- Community
- Author
- brnyxx
- Last verified
- Oct 2026
About this project
English · 한국어 · 日本語 · 简体中文. Site: brnyxx.github.io/jev-ra replays a real recorded run and explains the pipeline. A fast browser-use layer for CLI coding agents. Claude Code, Codex, or any MCP client hands jev-ra a goal. TypeSafe Jev, a System One decision model, picks the operation and the target element for every step in one round trip. Your agent plans, supplies the text values, reads what the page says, and takes over when jev-ra escalates. No second LLM runs inside the loop.
Quoted from the project's README. Read the full README.
Used inJev for Software & DevTools
- Languages
- Python
Highlights
- Reported 2,714 ms to open the Gödel article against 23,058 ms for browser-use 0.13.10 with gemini-3-flash, an 8.5x gap.
- One decision per step, and the only text typed into the page is text you supplied yourself as a value.
- 25 of 25 runs verified against the final page with no text-model calls; jev-ra figures are medians of 5 runs.
- Runs straight from PyPI through uvx with no install step, registering uvx jev-ra mcp as the server command.
Watch out
The speed comparison is one machine and one dedicated Chrome, both tools through OpenRouter, and the browser-use side is a single recorded run that was faster than its own five-run median on every task, so treat the ratios as a floor rather than a clean win. The container image may relaunch Chrome without its sandbox. 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.
The README reports 2,714 ms against 23,058 ms for browser-use 0.13.10 with gemini-3-flash on the Gödel article, and 3.96x on a Korean shopping sort. Both ran through OpenRouter on one machine and one dedicated Chrome, but jev-ra is a median of 5 runs while browser-use is a single recorded run that beat its own median on every task.
No. There is one decision per step, and the only text typed into the page is text you supplied as a value binding. jev-ra supplies the operation and the target element; your agent plans, supplies the text values, reads what the page says, and takes over when a run escalates with needs_human, which resume can carry on from.
Only on the container path. Locally it drives your own Chrome over the DevTools Protocol, and doctor checks the key, the endpoint, Chrome and one live decision before you start. The repository Dockerfile builds an image with a Chromium in it, and because Chrome's own sandbox needs a user namespace a container's seccomp profile usually refuses, jev-ra launches it once, reads what it said and starts it again without the sandbox when that is the cause.
No second LLM runs inside the loop; Jev is asked once per step for the operation and the target element. Every MCP response carries elapsed_ms, plus decisions and cost whenever Jev was called, and each run is stored under its own id so jev-ra profile RUN_ID can print where the time went step by step.
Refs go stale on purpose. browser_observe returns a page_key, and you pass it back with browser_click, browser_type or browser_select. A ref taken from a page that has changed since that observation is refused as stale rather than acted on, which stops a decision made against one DOM from being applied to a different one.
Give the run a profile. --profile NAME creates its own Chrome user-data-dir under $XDG_STATE_HOME/jev-ra/chrome-profiles/NAME, so a login done once is still there next time; open records the profile and the stateful commands reattach to it, and the MCP tool takes the same thing as browser_open(url, profile). One process drives one browser, so a server already on a profile refuses a second one.
serve exposes the same tool set over MCP's streamable HTTP transport at /mcp with server-sent events, and /healthz answers without a key. Every request carries a key from JEV_RA_SERVE_KEYS as an Authorization: Bearer header, is charged the decisions it spends against that key's daily quota, and returns its run id in the x-jev-ra-run-id header. The key is never logged; stderr names it only by a digest.
More Jev resources like this
19GitHub stars
MCP server where one browser_goal call hands a URL, a goal, and code-checked assertions to a server-side loop driven by a decision model over live page elements.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
jiawei686/jev-ultrafast-mcpPython
jiawei686/jev-ultrafast-mcpPython22GitHub stars
Independent open-source local decision model inspired by Jev, with a spark-s1 family plus smaller control models for agent gates, served locally with probabilities and policy actions.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
abhishek085/open-spark-jevPython
abhishek085/open-spark-jevPython114GitHub 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/fastbrowsePython1GitHub stars
A Python search command for terminals and agents: Parallel Search seeds queries, a local Chromium fetches pages concurrently, and Jev scores the evidence to expand a breadth-first link queue. Ships readable output, JSON and an MCP server.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
komikat/psearchPython
komikat/psearchPython2GitHub stars
Three dependency-free files that put Jev in a browser agent's hot path: code builds the choices from a table of visible controls, Jev picks an operation and target in one forward pass, and the code executes the pick.
Jev Demos & Experiments#community#jevbest#catalog-import
Community
zurfyx/jev-browser-skillJavaScript
zurfyx/jev-browser-skillJavaScript1GitHub stars
An open Apache-2.0 2B non-autoregressive decision model for agent action routing on a Qwen3.5-2B backbone: one forward pass over a candidate set with a pointer head and an escalation head, shipped as BF16, FP8 and INT8 weights.
Jev Demos & Experiments#community#awesomejev#catalog-import
Community
mkeco/Cerebellum-2BPython
mkeco/Cerebellum-2BPythonFrom 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.