Question 1

Is this another loop that screenshots its way through a task?

Accepted Answer

No, that is the thing it is built to avoid. Simple browser searches and URL navigation are handled by deterministic Windows code and consume zero TypeSafe tokens. General desktop tasks go through Microsoft UI Automation to produce a bounded action list that Jev then chooses among, one request per step, and the host LLM waits for the single MCP call and receives one final compact trace rather than spending a turn on every click.

Question 2

What will it refuse to do?

Accepted Answer

MCP mode blocks passwords, CAPTCHAs, elevation, financial transactions, destructive actions, sending messages or email, and publishing. Password fields never enter observations at all. Low-confidence decisions, unsupported completion and stalled interfaces stop the run, general tasks stop after 12 actions by default with a ceiling of 20, and it cannot control a process running at a higher Windows integrity level.

Question 3

Can I trust the research output without reading every page?

Accepted Answer

There is a trail to check against. Each completed step writes an evidence sidecar where every passage has a stable content-derived ID, canonical source URL, deterministic quality score, query and title, with duplicate URLs, tracking parameters and duplicate passages removed before storage. Claim verification is tied to one exact stored passage and routes confidence below 0.80 to review, and page contents are never sent to Jev at all.

Question 4

What does installation and configuration actually involve?

Accepted Answer

An install script with an optional Claude Code registration flag, an updater that requires an expected SHA-256 before it replaces installed files, and the guarantee that the installer never copies .env files. Put TYPESAFE_API_KEY in a .env.local beside the app or in a parent folder; the URL and search fast paths do not need it. Register the stdio server with claude mcp add or codex mcp add, and use the desktop setting or a config.toml entry on Codex CLI builds without mcp add.

Question 5

How do I install it and add it to my agent?

Accepted Answer

Run .\install.ps1 -RegisterClaude, or add the stdio server manually with claude mcp add windows-jev --scope user -- "C:\path	o\WindowsJev.exe" --mcp, or codex mcp add windows-jev -- "C:\path	o\WindowsJev.exe" --mcp. Restart the host after adding it. URL and search fast paths do not require a key.

Question 6

What does the research worker do?

Accepted Answer

start_research_job accepts up to 100 ordered inputs and returns a job ID immediately while an independent hidden worker downloads bounded public pages, extracts readable text and writes a cited Markdown artifact after every query or URL. Completed work is checkpointed and preserved across retry, and each step writes a .evidence.jsonl sidecar with stable content-derived IDs.

Question 7

What are its current limits?

Accepted Answer

UI Automation works best with standard Windows apps and accessible web controls; canvas, WebGL, games, remote desktops and custom-drawn controls may expose little state. This build is unsigned and lacks an installer UI, automatic updater and code-signing pipeline. Inaccessible controls, anti-bot pages, authentication, network outages and higher-integrity processes remain explicit failure states, recorded rather than silently claimed.

WindowsJev

About this project

Highlights

Questions about this entry

More Jev resources like this

From the Jev guides

From the Jev community