michaelswissa / jevry

Give your browser a task. Watch it work. 
 Browse, research, and work through supported websites in plain language. Jevry is an open-source desktop browser with visible actions, page-grounded results, and controls to stop or redirect the task. 
**▶ Watch the full Jevry introduction · 72 seconds**
Lightweight introduction excerpt · 8 seconds, 955 KiB · Media provenance
 Install ·
 Demo ·
 Why Jev? ·
 Architecture ·
 Benchmarks ·
 Contribute 
**Keep your tabs and conversations in one workspace.** Ask Jevry to find documentation, compare sources, or work through supported page controls, then follow its actions and check the result. Website support varies; the examples and measured results below show the current scope.
**Language models plan. Jev decides. Chromium acts.** TypeSafe’s Jev chooses among actions the browser can execute; a separate text model handles planning, language, and occasional visual reasoning. The desktop runtime validates and executes actions and retains evidence.
**Before you start:** this is a source preview, requiring Node.js 22.12+, a TypeSafe Jev API key, and a separate text-model connection. Model usage can cost money. Install · First-success guide
## See it in action
*Latest desktop design, captured from the local September 24 build.*
 Watch Jevry win 2048 — recorded with the earlier interface 
**“Play this game and win it.”** The page reaches a **2048 tile**, displays **You Win**, and reports **20,708 points in 1,013 moves, with no powerups used**.
Watch the 30-second real-time clip · Recording and edit details
The clip is an uninterrupted 1× excerpt from a September 23 run and shows the earlier interface. The GIF is a shorter crop focused on the game. The footage is preserved as recorded evidence and is separate from the instrumented latency study below.
## Why Jev?
Jev is TypeSafe's **System One model**: it evaluates supplied context and returns typed judgments and probabilities. Its interface exposes **Choice** for selecting an option, **Score** for rubric-based ratings, and **Noul** for yes/no judgments. Jevry's action loop is built around **Choice**.
That maps naturally to a browser. The page already contains a finite set of observed controls. Jevry turns those controls into choices such as `CLICK:17`, `TYPE_TEXT:4`, `SCROLL_DOWN`, `DONE`, and `BLOCKED`, then resolves the selected choice to an application-owned operation.
| Component | Responsibility in Jevry |
| --- | --- |
| **Jev** | Choose the next supported action; select observed targets and supplied field values; assess offered completion evidence. |
| **Text / vision model** | Interpret conversation, plan richer tasks, generate missing text, synthesize research, and perform visual setup or bounded recovery. |
| **Deterministic code** | Construct the action space, validate responses, check page freshness, dispatch input, retain receipts, and enforce stop conditions. |
**Jev chooses from the browser's actual capabilities.** Its output never becomes arbitrary J