v0.4.0: Jev can look

Jev can look now

Jev 1.13 reads images and Jev 1.14 adds a question type that boxes one thing in a picture. reflex-cu 0.4 uses both through a new look: true option:

• find with look finds icons and things in game scenes that have no text: a rough box on the whole window, then a second box on an enlarged crop around it. confidence is Jev's probability, after zooming in, that the target is really there; it clicks only when that is high enough.
• check / wait with look give Jev the window's screenshot along with the text read from it.
• The screenshot goes to Jev only, never into the agent's context.

Measured without clicking, against known positions: 8 of 8 targets on a game screenshot; 6 of 8 small browser-toolbar buttons (one refused for low confidence, one landed on a tab's title instead of its close button). A find takes about 1–3 s, a check about 0.6 s.

Limits. Experimental at TypeSafe ("Dreamer-ready"). Only a TypeSafe key works for now: OpenRouter already serves Jev 1.14 but drops the image, so look fails with an error there instead of answering blind. status reports jev_images. Boxing uses jev-preview until jev-latest moves to 1.14 (CU_JEV_VISION_MODEL to change).

---

Jev 1.13 起能看图，1.14 加了框选题型。0.4 新增 look: true：find 能找图标和游戏画面里的东西（整窗粗框 + 放大细框），check/wait 能连截图一起判断；截图只发给 Jev，不进代理的上下文。目前只支持 TypeSafe 官方密钥。

reflex-cu 0.3.5: computer use MCP server with Jev

reflex-cu is an open-source computer use MCP server for Claude Code, Codex and other AI agents. It controls macOS and Windows desktops, and a remote Windows machine over SSH, by reading the accessibility / UI Automation tree instead of taking a screenshot at every step. Jev, TypeSafe's System One decision model, answers the small decisions between steps.

What is in this release

• Three ways to reach Jev: a key from TypeSafe, OpenRouter or Command Code. On a network that cannot reach TypeSafe directly, an OpenRouter key works without a proxy in our tests.
• Control state, not just names (Windows and macOS): what a dropdown is set to and whether it is open, whether a switch is on, what a text field contains, which group a repeated button belongs to.
• Items scrolled out of view are listed on Windows and scrolled to automatically when clicked.
• macOS system panels such as Spotlight count as the foreground window, so waiting for them no longer times out.
• Batched steps report per-step timing and stop as soon as the judge is unsure.

Measurements

Same agent and task sentence, all tools versus the same server limited to screenshots and clicks, two runs each:

Task	All tools	Screenshots and clicks only
macOS: "open the ChatGPT app"	8 / 9 s, 4 calls	18 / 30 s, 7 / 12 calls
macOS: "open Calculator and work out 1234 times 56"	23 / 22 s	35 / 38 s
Remote Windows: five changes in Settings	61 / 88 s, about $0.27	54 / 72 s, $0.40–0.56

Install

Tell your coding agent: "Install reflex-cu for me, following https://github.com/8itlab/reflex-cu/blob/main/INSTALL_FOR_AI.en.md"

中文说明见 README。