JevChat Windows
Windows chat overlay: offline OCR, Jev judge on intent and tension, three ranked drafts, fill-only never auto-send.
Overview
JevChat Windows (github.com/jev-chat/jev-chat-windows) ports the jev-chat kernel to PyQt on Windows: window screenshot plus offline RapidOCR builds thread state, Jev judges intent, tension, and reply urgency, then a separate draft LLM writes three candidates ranked by Jev probabilities. Fill-in only; send stays manual. Releases ship as about 146 MB zip with jev-chat-windows.exe; keys for judge (OpenRouter or TypeSafe) and draft (default DeepSeek) live in HKCU environment variables. README documents session following, group speaker names, pause toggle, and optional debug overlay for OCR boxes.
Problem: Windows desktop chat users want the same judge-first reply assist as mobile, without hooking WeChat or auto-sending messages.
Built for: Windows 10 1903+ users who run chat apps locally and want OCR read, Jev intent scoring, and three ranked drafts they paste manually.
First indexed on Jev Directory: 2026-09-24
Creator and team
- Name
- jev-chat
- Organization
- jev-chat
How Jev is used
- Role in the product flow
- Judge pass on OCR thread text before draft LLM writes three ranked replies
- Primitives
- ChoiceScoreNoul
- State in
- Recent OCR messages per session, relationship preset, optional group reply target, and style notes from config.json beside the exe.
- Decision out
- Intent summary, tension 0 to 9, three reply candidates with Jev probability ordering; fill input only.
- WGC captures chat window; RapidOCR extracts lines in memory
- Judge API (Jev via OpenRouter or TypeSafe) scores intent and ranks drafts
- Draft LLM writes three candidates (default DeepSeek direct)
- User taps fill; program never triggers send
Windows shares the judge-then-draft split documented on jev-chat-jev-chat-jarvis for Android: structured judgment before casual text generation. Capture differs: desktop uses window screenshots and offline OCR instead of accessibility trees. TypeSafe or OpenRouter can back the judge card while draft defaults to DeepSeek for China-friendly latency per README. Cross-link the Android thick page for QQ and Lark paths; this listing covers WeChat-oriented Windows usage and floating PyQt UI. No package hooks and no auto-send mirror the mobile privacy story.
Sourced performance claims
- README states fill-in only with manual send and about 146 MB release zip size.Source: github.com/jev-chat/jev-chat-windows README
- Public GitHub repo jev-chat/jev-chat-windows had 530 stars and 115 forks when this listing was drafted.Source: GitHub API September 2026
- Judge and draft use separate API keys (JEV_API_KEY and LLM_API_KEY) stored in Windows user environment variables, not files.Source: github.com/jev-chat/jev-chat-windows README
Features and stack
Features
- Three ranked replies with Jev probability percentages
- Session-following overlay with pause toggle
- Separate judge and draft model cards and keys
- Group chat speaker OCR and optional reply target
- Debug view for OCR boxes without saving screenshots to disk in release exe
Stack
- Python
- PyQt
- RapidOCR
- OpenRouter or TypeSafe judge
- DeepSeek draft default
Pricing: Free open source client; judge and draft APIs bill per provider when messages arrive.
Links
FAQ
- Does the Windows app auto-send replies?
- No. README emphasizes manual send after fill-in, matching the Android jarvis policy.
- Which API keys do I need?
- One key for judge (OpenRouter or TypeSafe) and one for draft (default DeepSeek). They are stored separately in HKCU environment variables.
- How does this relate to Jev Chat Jarvis?
- Same jev-chat family kernel: judge-first Jev on visible thread text, three drafts, never auto-send. Jarvis is Android accessibility; this repo is Windows OCR plus PyQt.
Related learn guides
Original Jev guidance that pairs with this product pattern.
- Jev use cases
The patterns builders actually search for: moderation, routing, triage, RAG verify, and agent gates.
Related products
Hand-picked neighbors with rich profiles or overlapping tags.