Vertical Jev Bots: jev-chat-windows and HA-Jev

Most Jev projects on the Awesome Jev radar are developer infrastructure — routers, evals, browser agents. A smaller set builds directly for end users, wiring Jev's typed scores into a specific app most developers don't touch. Two of them land in very different corners of daily life: a Windows desktop companion for WeChat, and a Home Assistant integration for smart-home sensors.

Both are instructive for the same reason. Each one draws a hard, explicit line around what Jev is allowed to decide, and what a human or a hardcoded rule still controls. That line is worth reading closely before adopting either.

Key Takeaways

• jev-chat-windows never auto-sends messages. It reads WeChat windows with fully offline OCR, drafts three reply candidates with Jev, and requires a manual click to insert (not send) the chosen text.
• HA-Jev turns natural-language questions into typed Home Assistant sensors — probabilities, choices with distributions, or scaled scores — usable directly in automations.
• The HA-Jev author explicitly warns against safety-critical uses: locks, heaters and smoke alarms are called out by name as things Jev shouldn't gate.
• jev-chat-windows keeps API keys in the Windows registry, not in files, and the OCR step happens fully offline before anything reaches an API.
• Both are single-purpose tools with real star counts — jev-chat-windows at 398 stars, HA-Jev at 53 — but neither has been independently load-tested by this site.

jev-chat-windows: reads your screen, drafts replies, never sends

jev-chat-windows (listing) is a Windows companion for WeChat's desktop client (4.x). It captures screenshots of the WeChat window, extracts text with the fully offline RapidOCR engine — no data leaves the machine during OCR — and then runs a three-stage pipeline: Jev answers seven judgment questions about the message's intent and emotion, a separate LLM generates three candidate replies based on that analysis, and Jev ranks the three by probability. The chosen reply fills the WeChat input box with one click. The author states plainly, in Chinese in the README, "发送永远手动" — sending is always manual.

The judgment step (intent, emotion) runs through OpenRouter or a direct TypeSafe connection using a JEV_API_KEY; draft generation defaults to DeepSeek's API but supports 11 providers including OpenAI, Anthropic and Gemini via a separate LLM_API_KEY. Both keys are stored only in the Windows registry (HKCU\Environment), not in project files. The README is specific about what it doesn't do: no process injection, no memory reading, no access to the WeChat database. Known rough edges include a yellow capture frame that's unavoidable on Windows 10 (hideable on 11), OCR failure if the message input box exceeds 45% of the panel height, and occasional missed sender attribution in group chats. It requires Windows 10 1903+ or 11, WeChat Windows 4.x, and either the ~146MB standalone executable or Python 3.10–3.12 from source. Project code is MIT, but a bundled UI component (PySide6-Fluent-Widgets) is GPLv3 and requires a separate commercial license for commercial use — worth checking before shipping this inside a paid product. It has 398 stars as of September 2026.

HA-Jev: "ask your house a question, get a number back"

HA-Jev (listing) integrates Jev into Home Assistant with that line as its own tagline: ask a natural-language question about your home, get back a probability, a choice with a distribution, or a scaled score — not a chat response. Questions become sensors, readable in the dashboard or referenced in YAML automations. Four automation actions (jev.noul, jev.choice, jev.score, jev.ask) return response variables usable directly in existing automation logic, and there's an AI Task entity for voice-driven questions through Home Assistant's Assist, plus conversation-agent routing for spoken commands. Cost tracking is built in: the integration surfaces API call counts, token usage and estimated daily spend against a configurable token budget.

The example the project leads with — checking whether finished laundry has been sitting unattended — is representative of the whole approach: read existing entity states, ask Jev a typed question about them, let a configured threshold trigger an existing automation. The README ships 15 worked examples, four of which pair Jev with a separate LLM for more open-ended tasks. It's built on a jevclient library, requires Home Assistant 2026.9+ and a typesafe.ai API key (OpenRouter keys are also supported), and installs via HACS or manually. The one limitation stated without hedging: the author advises against using Jev for safety-critical controls — locks, heaters, smoke alarms are named explicitly. That's the right instinct for a probabilistic model making home-automation decisions, and it's a standard worth holding every Jev integration to, not just this one. MIT-licensed, 53 stars as of September 2026; the README references a fuller limitations page but this site hasn't independently verified latency or judgment accuracy beyond what's documented.

The shared pattern: narrow scope, explicit boundaries

Neither project tries to make Jev the decision-maker for anything consequential. jev-chat-windows keeps a human in the loop on every send. HA-Jev keeps Jev out of anything that could hurt someone if it's wrong. That's a healthier default than what shows up in some other corners of the ecosystem — see How to Evaluate Jev Projects Honestly for a fuller checklist — and it's worth treating as the bar any new vertical Jev bot should clear before you install it.

What to do next

• If you use jev-chat-windows commercially, check the GPLv3 terms on the bundled Fluent Widgets component before shipping it as part of a paid product.
• In HA-Jev, keep Jev-driven automations away from safety-critical devices, exactly as the author recommends — locks, heating, smoke alarms.
• Set a token budget in HA-Jev before enabling it broadly across a home's sensors; costs scale with how many questions you ask and how often.
• Read Wiring Jev into Agents with MCP if you want a more general pattern for routing typed decisions into existing systems, beyond these two specific integrations.