jev-chat-jarvis-mac
macOS chat overlay that reads the current chat window and shows the incoming message's intent, risk and candidate replies, read-only.
Overview
jev-chat-jarvis for macOS, documented in Chinese, is a floating overlay that, when a chat message arrives, shows its real intent, a risk level and how to reply. It is read-only and non-invasive: no injection, hooking or database decryption, only reading the screen. Supported chat apps are read either by window screenshot plus Apple Vision OCR, with bubble-based classification that separates message text, nicknames, pinned notices and quoted messages, or by reading system text interfaces directly with accessibility permission.
The judgment layer reports 86.4% zero-shot accuracy across 8 intent types on a 22-item regression set, a 0-9 risk score and an action suggestion, computed in one forward pass by a local model by default, with Jev, OpenAI-compatible and Anthropic-compatible keys configurable in settings. Fourteen built-in reply styles generate candidates concurrently, which are ranked and can be copied or filled in; sending is always manual. On an M1 Pro, intent appears in about 1.5 s and candidates in 1.5-2 s. It installs via Homebrew or a release app, and is MIT licensed.
What the decision model decides
| Type | Question | Answers |
|---|---|---|
| Choice | What is the intent of this incoming message (8 types)? | — |
| Score | What is the risk level (0-9)? | — |
| Choice | Which candidate reply ranks highest? | — |
How it works
- Detect a new message via OCR or accessibility text
- Judge intent, risk and action
- Generate candidates in selected styles
- Rank candidates; user copies or fills one
- User sends manually
Strengths
- Read-only; no injection or database access
- Fast local judgment by default
- Detailed message and quote attribution
Caveats
- Default judgment uses a local model; Jev is optional
- Small regression set for the accuracy figure
- Requires screen recording and accessibility permissions
Getting started
`brew tap jev-chat/tap && brew install --cask jev-chat/tap/jev-jarvis`, then grant screen recording permission.At a glance
- Type
- Project
- Status
- beta
- Language
- Python
- Runtime
- macOS overlay app
- Has code
- yes
- Has tests
- yes
- License
- MIT
Use cases and integrations
How curated lists describe it
macOS chat overlay copilot that reads the current conversation and calls Jev to judge intent, risk, and reply suggestions.
logicrw/awesome-jev-projects聊天悬浮窗助手（macOS）：屏幕感知 + 本地小模型判断意图与风险，按话术生成回复候选。纯只读.
wh000wh000/awesome-jev-live微信消息意图识别悬浮窗（macOS）：看屏 + 本地小模型判断意图和风险，再按话术生成回复候选。纯只读、不注入微信。
hellogumbo/awesome-jev
Frequently asked questions
Does it use Jev?
Jev is one configurable judgment backend; a local model is used by default.
Does it modify the chat app?
No, it only reads the screen or system text interfaces.
How accurate is intent detection?
86.4% on a 22-item regression set, per the README.
Related projects
- jev-chat-windows★ 802
Windows WeChat reply assistant: local offline OCR reads the chat, Jev judges intent and tension and ranks three drafted replies to fill in.
ProjectproductionPython AstrBot plugin where Jev decides whether the bot should join a group conversation; the main model still writes the reply.
ProjectproductionPythoniOS custom keyboard that shows a copied message's intent, risk level and candidate replies inside any chat app, filling but never sending.
AppbetaSwiftWindows desktop chat assistant that OCRs a selected chat region, has Jev analyse the conversation, and ranks drafted replies for copying.
ProjectbetaPython, TypeScriptUnofficial fork of jev-chat-windows: WeChat overlay with local OCR and drafted replies, adding optional auto-send and a knowledge base.
HookbetaPythonWeChat reply assistant fork: DeepSeek drafts three replies, Jev pre-judges, reviews and ranks them; auto-send is cancellable.
ProjectbetaPython