What is this?
A floating overlay for WeChat on macOS: the moment a message arrives, it shows the sender's likely intent, a 0-9 risk score with a suggested action, and ready-to-use reply candidates.
A floating overlay for WeChat on macOS: the moment a message arrives, it shows the sender's likely intent, a 0-9 risk score with a suggested action, and ready-to-use reply candidates.Is it a WeChat plugin or bot, then?No — it is a standalone screen-watching app: it screenshots the WeChat window, OCRs the chat bubbles and judges them with models. It never injects, hooks or decrypts anything, which the README bills as zero ban risk.
No — it is a standalone screen-watching app: it screenshots the WeChat window, OCRs the chat bubbles and judges them with models. It never injects, hooks or decrypts anything, which the README bills as zero ban risk.How do I install it on my Mac?Download the .app from GitHub Releases, unzip it into Applications and right-click → Open the first time — it is unsigned, so a double-click gets blocked by Gatekeeper. Then grant Screen Recording permission on first launch; if macOS calls the app 'damaged', a Terminal command clears the quarantine flag.
Download the .app from GitHub Releases, unzip it into Applications and right-click → Open the first time — it is unsigned, so a double-click gets blocked by Gatekeeper. Then grant Screen Recording permission on first launch; if macOS calls the app 'damaged', a Terminal command clears the quarantine flag.Do I need my own API key, or can it run fully offline?Both are optional. The judgment layer takes either a cloud key or a local offline model (~3.8 GB, offered on first launch); the reply layer ships with a built-in shared key in the packaged app. Pair the local model with Ollama and nothing leaves your Mac.
Both are optional. The judgment layer takes either a cloud key or a local offline model (~3.8 GB, offered on first launch); the reply layer ships with a built-in shared key in the packaged app. Pair the local model with Ollama and nothing leaves your Mac.What are its weak spots?The overlay only appears and analyzes while WeChat is in the foreground. Images and stickers can't be read, quoted replies are treated as plain text, and a WeChat redesign can break the layout constants until recalibrated. Messages whose sender side is unclear are skipped.
The overlay only appears and analyzes while WeChat is in the foreground. Images and stickers can't be read, quoted replies are treated as plain text, and a WeChat redesign can break the layout constants until recalibrated. Messages whose sender side is unclear are skipped.Will it run on any Mac?Only Apple Silicon (M-series) Macs on macOS 13+. Intel is out, and Rosetta will not help either — the local judgment model needs torch, which has no Intel build; the app detects this at startup and tells you.
Only Apple Silicon (M-series) Macs on macOS 13+. Intel is out, and Rosetta will not help either — the local judgment model needs torch, which has no Intel build; the app detects this at startup and tells you.What do the reply candidates look like?Eleven built-in tones generate candidates in parallel — three are on by default (high-EQ, Tieba veteran, passive-aggressive), each producing one steady and one bold variant, re-ranked by the local model. You copy a candidate or fill it into the input box; sending always stays manual in WeChat. Custom tones can be added via a JEVTONES env line.
Eleven built-in tones generate candidates in parallel — three are on by default (high-EQ, Tieba veteran, passive-aggressive), each producing one steady and one bold variant, re-ranked by the local model. You copy a candidate or fill it into the input box; sending always stays manual in WeChat. Custom tones can be added via a JEVTONES env line.How fast does it react to a new message?Intent judgment and candidate generation start at the same time. On an M1 Pro the intent lands in about 1.5 s and candidates in 1.5-2 s; the very first judgment after launch is slower while the local model warms up.
Intent judgment and candidate generation start at the same time. On an M1 Pro the intent lands in about 1.5 s and candidates in 1.5-2 s; the very first judgment after launch is slower while the local model warms up.Which model providers are known to work?The README names three tested setups: DeepSeek's deepseek-chat (the fastest), Zhipu glm-4-flash via the Anthropic-compatible settings, and a local Ollama qwen2.5:7b for zero outbound traffic. Avoid 'thinking' models — their reasoning eats the token budget and you end up with zero candidates.
The README names three tested setups: DeepSeek's deepseek-chat (the fastest), Zhipu glm-4-flash via the Anthropic-compatible settings, and a local Ollama qwen2.5:7b for zero outbound traffic. Avoid 'thinking' models — their reasoning eats the token budget and you end up with zero candidates.Does it remember earlier messages for context?Only if you turn chat history on in settings (it is off by default). It then keeps the last 100 actually-read messages per chat in a local file, never touching the WeChat database, and you can attach free-form background notes like 'AAA is the group owner' to help the model.
Only if you turn chat history on in settings (it is off by default). It then keeps the last 100 actually-read messages per chat in a local file, never touching the WeChat database, and you can attach free-form background notes like 'AAA is the group owner' to help the model.What are similar projects?DigUp, lipflow, Baize, WechatVibe, Kun, SCM, Minke, open-codesign
DigUp, lipflow, Baize, WechatVibe, Kun, SCM, Minke, open-codesignWhich topics is it in?anthropic-claude, screen-capture, on-device-ai, on-device, ai-assistant, wechat, ocr, macos-app, openai-compatible, local-llm, apple-silicon, ollama
anthropic-claude, screen-capture, on-device-ai, on-device, ai-assistant, wechat, ocr, macos-app, openai-compatible, local-llm, apple-silicon, ollama