1. The Production Bottleneck
Browser workflows expose many candidate links, buttons and inputs, while page state changes after every action. A generative controller can mix planning, element selection and execution in one opaque loop, making stale targets and recovery behavior hard to inspect.
2. Architecture Comparison: Without vs With Jev
Planning, choosing a target and executing it are difficult to measure separately.
The controller may resend more page state than the next bounded choice requires.
Stale elements, failed navigation and task completion can be mistaken for progress.
Serialize the current action space and let application code retain the selected target.
Make completion, retry and human-review outcomes part of the contract.
The browser layer still owns permissions, navigation, waits and side effects.
3. Execution Micro-Flow
How state flows through the bounded decision layer during live agent execution:
4. Practical Engineering Trade-offs
✓Deploy This Pattern When:
- The next action comes from a known candidate set
- The browser can expose structured accessibility or DOM state
- The loop needs explicit stop, retry or escalation decisions
⚠Avoid Or Re-evaluate When:
- Visual-only targets and CAPTCHA flows
- Authentication, permissions and irreversible actions
- Untrusted page content being treated as instructions
5. Three browser-agent implementations
These projects divide the work differently. Their demos and figures are reported by their authors; this site has not benchmarked the three systems against each other.
Browser Use + Jev: indexed actions
Each page observation becomes a numbered table of available controls. Jev chooses an operation and a compatible target; a separate small model writes text only for typing actions. Browser code rechecks the target and verifies the outcome.
Best reference: a goal-driven browser loop with an explicit action space. The reported flight run is one example, not a general speed guarantee.
Watch the sourced flight case →Read the project repository ↗Stagehand + Jev: remote execution
Kyle Jeong describes a remote-browser demo that sends the accessibility tree and candidate actions to Jev, then lets Stagehand execute the selected step.
Best reference: the decision/execution split in a remote browser. The post does not provide a public implementation repository for this exact demo.
Watch the sourced Stagehand case →Read the original post ↗Jev Browser: MCP, CLI and library
Joey Kudish's project offers the same browser loop through three interfaces. Jev chooses a page action and answers goal/stuck questions; code owns time and step budgets, recovery and stopping.
Best reference: inspecting a step trace and the surrounding control flow. The repository labels the software early and warns of rough edges on harder sites.
Inspect the sourced project →Read the project repository ↗6. Sourced Community Builds
74 ImplementationsCapture a Screen Region by Describing It
Name the screen content you want and watch the selection move to it.
Vlad Terin Jev Browser
A Codex browser adapter in which Codex plans, Jev selects the next action and the browser executes it.
Jev Use - Mac Calendar Automation Benchmark
10x faster macOS computer use with Codex and Jev eliminating LLM thinking pauses.
Jev Browser
A Jev-driven browser agent exposed as an MCP server, CLI, or library.
Wikiracing Fast Navigation
An official benchmark demonstrating Jev selecting optimal Wikipedia links in real time to reach target concepts.
CUA-S1 Form Automation
An open 706k-parameter System 1 model specialized for mapping user data into web form inputs in 50ms.
Jev-RA Reactive Browser Agent
A browser-use layer where Jev chooses the operation and target element for coding agents.
Flight Search with Browser Use
A browser agent searching real-world flights at 1× speed with Jev choosing each next action.
jev-browser in Cline
A marketplace plugin that connects Jev-driven browser execution directly into the Cline coding agent.
Computer Use Without Screenshots
On-device CoreML segmentation and OCR feeding text tokens to Jev for privacy-first desktop control.
TypeSafe Computer Use
Open-source macOS desktop control loop with bounded Jev decisions at every single step.
AskJev (MCP Agent)
Claude handles conversational intent while Jev directs browser navigation across live websites.
Stagehand on a remote browser
we built blazing fast computer/browser use with Jev + @Stagehanddev.
Voice-controlled computer use on a Mac
I built a voice controlled computer-use for my mac using @typesafeai's Jev and it's INSANE how fast it is!
YouTube sponsor skipper
Just trying out Jev, I made a Chrome extension that:
wtf is jev? Demos: Browser Use, Classifier, Router
Demonstrates browser automation, model routing, and tokenless agent loops.
Jev and Browser Use, fully tested
Validates incoming actions against injection patterns before navigation.
Jev full tutorial: 3 prototypes built live
Scores relevance and routes between local voice input and browser execution.
Massively Parallel Adversarial Browser Testing
A high-concurrency headless browser adversarial test suite testing web releases with bounded Jev DOM navigation.
Smart Clipboard Resume-to-Form Autofill
Reinventing copy-paste by mapping unstructured resume text into complex multi-field hiring application forms instantly.
Jev Browser Skill (jb)
A standalone Playwright Chromium CLI that uses Jev to choose in-page browser actions.
0x_rody Jev Agent Workflow
A Jev install/try-it list, including repositories for browser control and context compaction.
Agentic browsing in Chrome
A Chrome extension where agents browse and click for you, with Jev deciding.
A chief of staff for bots
Jev reads the task, picks the agents and gives each one a model.
browser-ai
A Jev browser agent in pure Go, attached to a real Chrome.
Launching agents on a canvas
Orchestration without waiting for a slow LLM loop.
dsh-jev-tools
Judgment, not generation, inside DeepSeek Harness.
End-to-end tests run by agents
An open-source framework for agent-run e2e tests on web and mobile.
fastbrowse
Jev picks each action; an LLM plans; every claim cites the page.
Jev driving real Chrome from Grok Bot
Bots use Jev for the clicks instead of a slow look-and-click loop.
Headless Chromium agent
Jev drives headless Chromium so agents can browse quickly.
Interlock
A gate on every agent tool call, with Jev as one of its sensors.
One-click invoice finder
Finds the billing page on any website and lists every invoice.
jev-browser (Playwright)
One Jev and Playwright core behind a typed SDK, a CLI and an MCP server.
jev-browser, a general browser skill
Give it a site and a task, and Jev decides every click.
Jev Browser Use
A Codex skill: Jev clicks, Codex thinks and checks the result.
jev-browser
Browser automation where an LLM plans and Jev decides.
jev-cua
A local computer-use fast path for Codex and Waku.
jev-desktop
Jev picks the action inside Codex Computer Use.
jev-flash-router
An MCP server so coding agents stop spending tokens on small choices.
jev-harness
Every tool result is filtered by Jev before the model reads it.
jev-job-hunter
A job crawler that finds the Careers page and scores each role against your profile.
jev-mobile
Android control loops with Jev and Mobile MCP.
Jev Recruiter
A LinkedIn recruiting agent that browses profiles and saves the evidence.
jev-reflex
Claude thinks, Jev reacts: browser tasks at about 100 ms a decision.
jev-use
Say it, and your Mac does it, with no vision model.
jev-use (voice)
Voice control for macOS where Jev is the fallback, not the router.
jev-voice-browser
A browser you control by voice, from Moritz Kremb’s Jev tutorial.
jevcumber
Cucumber tests from the .feature file alone. No step definitions.
jevwire
A decision layer for agents: MCP server, library and Claude Code plugin.
LinkedIn NoSlop
A Chrome extension that blurs low-value LinkedIn posts.
Mobile Jev
A mobile agent on a real Android phone, with Jev making each decision.
A chat bot with no LLM
Jev picks the tool and its arguments, so answers are instant.
playjev
Typed browser automation: Playwright acts, Jev decides.
A second-hand shopping agent
Reads about 26 listings a minute and decides on each in 406 ms.
solari-reflex
A speed layer for computer use: one Jev decision per step.
super-jev
A small decision-to-action harness for Jev.
TipTour
A pointer companion for the Mac, with Jev driving the cursor.
eve
Vercel’s open agent framework, with Jev as its default evaluate model.
A computer assistant that listens
Local Whisper hears everything; Jev decides which action to take, in real time.
X timeline labeler
A Chrome extension that labels every post on your X timeline.
Agent Desktop - Reliable OS Accessibility Automation in Rust
Desktop computer use via OS accessibility trees with Jev choosing stable UI targets.
typesafe-mario - Direct Emulator RAM State Control with Jev
Plays Super Mario Bros. by reading structured emulator RAM without computer vision.
OneVOneJev - 1v1 Browser FPS Arena Powered by Jev
Fast-paced Three.js quickscope arena with Jev making real-time tactical decisions.
jevpilot - Three.js Driving Simulator with Jev Autopilot
Playable Three.js driving simulator with Jev selecting route options and speed regimes.
JevHarness
LLM-written Jev harnesses, with reward reflection over the full trajectory.
ego-jev
One ~0.4 s typed decision per DOM step, instead of an LLM turn.
dsh-jev-prune
Context compaction for DeepSeek Harness, judged by Jev.
Agent Chaperone
Screen an agent's tool calls with Jev before they run.
JevLoop
An agent loop where the decisions cost no LLM call.
Stencil QA Agent
A QA loop where Jev selects browser actions while a frontier model plans and vision verifies.
jev-browser-wingman
A bounded browser-action tool for Claude Code and OpenCode on a real logged-in Chrome session.
jev-browse
One-call browser sub-tasks for Claude Code and Codex, with Jev choosing each low-level action.
jev-kit for Claude Code
A Claude Code kit that wires Jev into guards, sub-agent sizing, search, browsing and review.