jev-chat-jarvis (macOS)

Neither. The screenshot path grabs the chat window and runs Apple Vision OCR, and a second path reads the system text interface directly with no screenshot at all. There is no injection, no hook, and no database decryption, and sending a message stays a manual action in the chat app.

No. Only Apple Silicon M-series Macs on macOS 13 or newer are supported, and Rosetta is not a workaround because the local judgment model depends on a torch build with no Intel version. The app checks this at launch and tells you why it will not continue.

It is zero-shot accuracy across 8 intent classes on a 22-message Chinese regression set, and the repo ships that regression as a runnable script so you can re-run it. The latency figures of roughly 1.5 s for intent and 1.5 to 2 s for candidates on an M1 Pro are labelled in the README as a mechanism-derived estimate, with real numbers in the logs.

With history recording off, which is the default, only the current message and any background text you wrote go to whichever model service you configured. Enabling history stores up to 100 messages per chat locally at 600 permissions, but that text is still sent to that service for inference, and the generation layer's bundled shared key means candidates go to that key's provider unless you replace it.

Download the .app from Releases, unzip it, drag it into Applications, and right-click Open the first time because it is not notarized. If it reports the app is damaged, clear the quarantine attribute with sudo xattr -r -d com.apple.quarantine /Applications/jev-jarvis.app.

A judgment layer that can use TypeSafe Jev or a local decider-2b of about 3.8 GB, and a generation layer that accepts OpenAI-compatible or Anthropic-compatible endpoints. Both keys are optional: without a judgment key the first launch guides you to configure one or download the offline model, and without a generation key the bundled shared key drafts candidates.

Screen recording is needed for the screenshot-and-Vision-OCR path, and accessibility is needed both to read the system text interface and to fill a candidate into the input box. Apps using only the text-interface path need accessibility alone, and you must quit and reopen after granting screen recording.