jev-ultrafast-mcp
MCP server that takes a whole browser task (URL, goal, checks) in one call and drives Chrome server-side with a decision model.
Overview
jev-ultrafast-mcp lets a coding agent hand off browser work instead of driving the page click by click. The agent sends the URL, the goal and the checks that prove success in one browser_goal call; the loop runs server-side, a decision model picks targets from ref-based element tables of what the page actually has, and code-checked assertions verify the result. The server refuses rather than guesses when nothing fits. Simply looking at a page uses browser_open and the element table without any model call, and successful flows can be replayed as macros with zero model calls.
It speaks the Chrome DevTools Protocol to any Chromium-family browser. An install script detects WorkBuddy, Claude Code, Claude Desktop, Codex CLI, Cursor, VS Code, Cline, Windsurf and Gemini CLI and merges the right config into each, backing up the original. The README frames the economics as one tool call instead of twenty, seconds instead of a minute, and about a cent for a day's use. It has CI and requires Python 3.10+.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | Which observed element and operation advance the browser goal?(inferred) | — |
How it works
- Agent calls browser_goal with URL, goal and verification checks
- Server opens the page and builds a ref-based element table
- Decision model picks actions in a loop; refuses if nothing fits
- Code checks assertions; successful flows can be replayed without a model
Strengths
- One tool call replaces many agent turns
- Never invents targets
- Zero-model macro replay
- Installer covers many MCP clients
Caveats
- Requires a Chromium browser and a decision-model key
- Cost and speed figures are the author's
Getting started
git clone, python3 -m venv .venv, .venv/bin/pip install -e ., python scripts/install.py, then restart your MCP client.At a glance
- Type
- MCP server
- Status
- beta
- Language
- Python
- Runtime
- MCP server (CDP)
- Has code
- yes
- Has tests
- yes
Use cases and integrations
How curated lists describe it
Hand the browser work off: an MCP server where a decision model drives the page for your agent, so a flow costs one tool call instead of a turn per click. Ref-based element tables, code-checked assertions, zero-model macro replay. Speaks CDP to your Chrome.
hellogumbo/awesome-jevMCP server that takes a whole browser task in one call and drives the page itself, so the agent never opens a browser.
MrJev/awesome-jev
Frequently asked questions
Why not let my agent drive the browser?
Handing the whole task to the server takes one tool call instead of a turn per click.
What if the page has no matching element?
The server refuses rather than guessing.
Which clients can it configure?
Claude Code, Claude Desktop, Codex CLI, Cursor, VS Code, Cline, Windsurf, Gemini CLI and WorkBuddy.
Related projects
Jev browser agent that pierces shadow DOM and iframes, adds canvas fallback and full actions, with one Jev request per cycle.
MCP serverbetaPythonDrop-in proxy for Playwright MCP that adds Jev page triage, prompt-injection masking, snapshot pruning and risky-action gating.
MCP serverbetaTypeScriptMCP server and libraries that drive Chrome over CDP and hand a decision model like Jev a bounded menu of page actions.
ProjectproductionTypeScriptMCP server giving coding agents fast typed decisions (ship gates, risk checks, file triage, page checks) as accept, review or escalate.
MCP serverbetaTypeScriptAgent-plugin browser tool where Jev picks one operation and one row from a numbered page table; a small LLM only writes typed text.
ProjectbetaTypeScriptChrome extension plus MCP server so Claude or Codex can drive your real Chrome, with Jev picking each click in about half a second.
MCP serverbetaTypeScript