JARVIS runs as a first-class citizen of macOS — not a browser tab, not a cloud dashboard. One assistant, every surface: Command Center HUD, menu bar orb, floating overlay, voice. Built on the Stark Industries visual language.
Calendar, Reminders, Contacts, iMessage, Files, and system stats — controlled through reverse-RPC, not screen-scraping. JARVIS is a resident of your Mac, with FaceGate approval on anything world-changing.
On-device wake word, streaming speech-to-text, and sentence-by-sentence TTS. Speak with the on-device cloned Jarvis voice (current builds), or pick Kokoro, system voice, ElevenLabs, or Google Cloud — voice is a pluggable engine, not a lock-in.
Identity core, SQLite-FTS recall, and sqlite-vec semantic search working together — so JARVIS remembers who you are, what you said, and what you meant.
Proactive daily briefings, an initiative pass that nudges you unprompted, and a nightly reflection cycle that rewrites what JARVIS has learned — all running while you work.
A Command Center HUD surfaces CPU, memory, connection state, and active tool calls in real time — the same telemetry JARVIS itself reasons over.
Fast, standard, and deep reasoning tiers routed per-turn, so trivial requests stay instant and hard problems get the full orchestrator.
Wake word and streaming speech-to-text run on-device. Or type from the HUD or overlay.
The local orchestrator on :8765 pulls memory, then picks a fast, standard, or deep model tier for the turn.
Calendar, Reminders, Contacts, iMessage, Files and more are called over reverse-RPC into the Mac app.
Anything world-changing waits for your Touch ID before it runs.
The answer is spoken sentence by sentence while it is still being written.
mac__spotify tool.Illustrative examples of the kinds of requests JARVIS handles through its native tools.
Who you are, your preferences, standing instructions. Loaded on every boot, never forgotten, hand-curated over time.
LOADED ON EVERY BOOTFull-text search over everything said and done — fast, exact, and queryable the moment it happens.
INDEXED AS IT HAPPENSEmbedding-based recall for meaning, not just keywords — JARVIS finds the relevant memory even when the words don't match.
MATCHES MEANING, NOT KEYWORDSA navigable 3D map of what JARVIS actually holds: memories wired together by meaning, skills and schedules, coding agents, and every tool. It lights up live as JARVIS recalls, writes, and calls tools.
Nothing is faked. A region with no data stays empty, and an event for a node that doesn't exist is dropped.
Drag to orbit · scroll to zoom · click a node to fly to it and read it
Simulated preview. The values below are generated in your browser to show what the in-app Command Center looks like — they are not live data from a real install.
jarvis-core, runs on your Mac at localhost:8765.Jarvis.v.1.0.0.zip from the latest release, unzip it, and move Jarvis to Applications. Running from /Applications avoids two known debug-build quirks.GEMINI_API_KEY in Settings, then choose a voice engine. An ANTHROPIC_API_KEY is only needed for the Claude backend.memory/identity.md. That is what makes it your assistant.⌘ for Quick Prompt, or say Hey Jarvis.
macOS 26+. Native Swift app, bundled local orchestrator, and a choice of voice engines. Bring a Gemini API key and one assistant is on every surface.
The app, wake word, speech-to-text, memory, and the Kokoro voice (plus the cloned Jarvis voice in current builds) run on your Mac. Reasoning calls Gemini (or Claude), so answers need a connection.
The app itself is free to download. Gemini or Claude usage is billed by those providers against your API key, and the on-device Jarvis and Kokoro voices are free. A per-turn trace log shows token and cost readouts.
Voice: the cloned Jarvis voice (on-device), Kokoro (local), system voice, ElevenLabs, or Google Cloud. The Jarvis voice clones itself from a bundled reference clip the first time you use it, and you can re-clone it from your own audio clip in Settings. Current builds: the cloned voice is the default in current source builds. The v1.0.0 release predates it and ships with Kokoro as the default, so build from source or wait for the next release to get it. Models: Gemini by default, routed per turn across fast, standard, and deep tiers. Set MODEL_BACKEND=claude to let Claude run the turn loop instead.
No. Sensitive actions are gated by FaceGate and need your Touch ID approval first.
An identity file you write, full-text recall over every turn, and semantic recall filled in passively in the background. A nightly reflection pass rewrites learned.md, and a memory timeline shows what JARVIS has learned. Everything is stored locally.
macOS 26 or later. iPhone and Apple Watch apps are built but still need on-device testing.
mac__ui)jarvis-coreFull notes for each release are on the GitHub releases page. Plans can change.