Jev Chat Jarvis
Android chat co-pilot overlay: Jev judges intent and danger on screen text, ranks three reply drafts, fill-only never auto-send.
Overview
jev-chat-jarvis (github.com/jev-chat/jev-chat-jarvis, chatjevs.com) is an Android 11+ accessibility overlay that reads only what is already on screen, runs a judge pass for intent, danger, and reply urgency, then drafts three ranked reply candidates you paste into the input box yourself. It never auto-sends and skips transfers or red packets. QQ and X are verified on device; Lark uses accessibility bounds plus on-device ML Kit OCR when message text is not in the AX tree. WeChat Android support stopped at 1.4 because newer builds hide message text from ordinary accessibility. Desktop siblings jev-chat-mac and jev-chat-windows share the same kernel with different capture paths.
Problem: Mobile chat apps bury you in threads where a blunt autocomplete draft is worse than no help, especially when tone, intent, and scam risk all matter at once.
Built for: Android users who want a co-pilot overlay on QQ, X DMs, or Lark without hooking chat apps or giving up send control.
First indexed on Jev Directory: 2026-09-23
Creator and team
- Name
- jev-chat
- Organization
- jev-chat
How Jev is used
- Role in the product flow
- Pre-reply judgment on visible thread text, then reply drafting gated on judge outputs
- Primitives
- ChoiceScoreNoul
- State in
- Recent on-screen messages, session title, optional local knowledge notes and contact profile hits from on-device storage (README analysis section).
- Decision out
- Intent summary, danger level on a 1 to 9 rubric, should-reply-now signal, best-action Choice, then three candidate replies ranked by fit Score.
- Accessibility or OCR capture builds a text state snapshot
- Judge model returns intent, danger, urgency, and action in about one second per README
- Reply model drafts three candidates; judge re-ranks with proportion scores
- User taps fill or copy; ACTION_SET_TEXT or clipboard only, never send
The product splits judgment from generation on purpose: the judge route can point at TypeSafe Jev, OpenRouter, or vendor presets such as the documented Bocha Jev endpoint, while reply and vision routes stay independently configurable. That keeps intent and danger nouls on a fast structured path before any chat model writes casual text. Local notes and contact aliases inject only when tags or titles match, so replies stay consistent with your own facts without shipping a full cloud memory product. The overlay model is the opposite of bot APIs: no package hooks, no database reads, only what you already see. jev-chat-jev-chat-windows is the thick listing for the PyQt OCR desktop port.
Sourced performance claims
- Judge pass documents about one second latency with confidence on intent and danger outputs.Source: github.com/jev-chat/jev-chat-jarvis README
- Public GitHub repo jev-chat/jev-chat-jarvis had 5214 stars and 1009 forks when this listing was drafted.Source: GitHub API September 2026
- Release APK jev-assistant-v1.4 targets Android 11 plus with signed builds in apk/ and GitHub Releases.Source: github.com/jev-chat/jev-chat-jarvis README
Features and stack
Features
- Floating panel with danger, intent, and three ranked replies
- Fill input only; no auto-send or payment actions
- QQ, X, and Lark paths documented with OCR fallback
- Local knowledge base and optional on-device chat history
- Separate judge, reply, and vision API cards with connectivity tests
Stack
- Android Kotlin
- Accessibility services
- ML Kit OCR
- TypeSafe or OpenRouter judge routes
Pricing: Open source client; you bring API keys for judge, reply, and vision providers per README.
Links
FAQ
- Does the app send messages for me?
- No. README states the tool only fills the input box or copies to clipboard and never triggers send, transfers, or red packets.
- Does it read WeChat on Android?
- WeChat support stopped at version 1.4 because recent WeChat builds hide message text from standard accessibility and may block screenshots.
- Where do judge probabilities live?
- Configure the judge card for TypeSafe-compatible endpoints or OpenRouter. Reply ranking uses the same judge route to sort three drafts.
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.