An open-source, local-first AI assistant that grows its own skills, remembers how it reasons, tells you when it's unsure, and lets you undo anything. It runs on your machine, for free.
Quickstart · Docs · Landing · Roadmap · Discussions
The jarvis name is attached mostly to voice clones that open a website and read the weather. H.O.T-Jarvis has a sharper goal: an assistant you can actually trust, that gets more capable the more you use it, with nothing to pay and nothing leaving your machine. It ships as a real desktop app (Tauri v2), not a web demo, and every claim below is backed by tested code (around 90 tests, CI-gated).
All four are built and tested, not roadmap items.
| Feature | What it means | |
|---|---|---|
| 🧩 | Self-evolving skill library | Ask for an ability and it writes the code and a test, proves the test passes, and refines on failure. Untested skills are flagged and refuse to run. |
| 🧠 | Reflective reasoning-memory | It re-reads its own action log, keeps short lessons about what worked and what failed, and applies them to future work. |
| 🎚️ | Calibrated confidence | Every answer carries a self-rated score. Below a threshold it asks a clarifying question instead of guessing. |
| ⏪ | Replay and undo | Every action is recorded and reversible, with an audit that proves the log reproduces memory exactly. |
Read the reasoning behind each in docs: the four hero features.
Three commands to a running assistant that remembers you across restarts:
ollama pull llama3.2 # 1. free local model (install from https://ollama.com)
cp .env.example .env # 2. optional: add a free Groq / OpenRouter key instead
npm install && npm run tauri dev # 3. launch the desktop appPrerequisites: Node.js 20+, Rust stable, and on Linux the Tauri system deps. Full walkthrough in the quickstart guide.
The desktop app is real and running. iOS is planned, and the honest tradeoff is inference: there is no Ollama on a phone, so mobile keeps the local-first promise by acting as a companion to your desktop (with an on-device model or a free cloud tier as options). The rest of the core carries over unchanged.
The full mobile plan, including the App Store readiness checklist, is in docs/ios.
Inference runs on your machine through Ollama by default, with free cloud tiers (Groq, OpenRouter :free) only as a fallback. No paid API is ever required.
| Provider | Cost | Notes |
|---|---|---|
| Ollama (local) | Free, unlimited | Default. Private, nothing leaves your machine. |
| Groq | Free tier | Fast. Key at console.groq.com |
OpenRouter :free |
Free tier | Many models. Key at openrouter.ai |
The router prefers the local model, caches identical requests, and backs off from any cloud provider that rate-limits, so the app never pressures you to pay. Your conversations, skills, and memory live in a local folder you control, and you can export all of it as one JSON file or wipe it at any time.
A Tauri v2 shell with a web UI over a Rust core. The rule is that all real logic lives in a Tauri-independent core, so every module is unit-tested without a webview.
src-tauri/src/core/
router local-first model routing + cache + backoff
memory SQLite (messages, facts, insights) + export/wipe
skills sandboxed Rhai skill engine (save, test, version, run, roll back)
authoring the assistant writing its own skills
reflection digests the event log into lessons
confidence the self-rating on every answer
eventlog append-only action log
replay rebuild-from-log + determinism audit
More in docs: architecture.
Released v0.1.0 with all four hero features, plus a Jarvis-style HUD, a landing page, and a documentation site. Voice works both ways now: replies through your OS voices, and dictation through a local Whisper model that never uploads your audio (see talk to Jarvis). It lives in the system tray with a global hotkey. Wake-word conversation and the autonomous work loop are next. See the roadmap.
The one hard rule: the assistant must stay free to run. Small pull requests with tests, green CI, conventional commits. See CONTRIBUTING.md, SECURITY.md, and CODE_OF_CONDUCT.md. Licensed under Apache-2.0.
The hero features draw on ideas from the Voyager skill library, MUSE-Autoskill, Reflexion, ReasoningBank, "Hindsight is 20/20", Mem0, "Agentic Uncertainty Reveals Agentic Overconfidence", and the replayable-agent literature. Implementations here are original; the ideas are credited in docs/DECISIONS.md.