A local-first, voice-capable developer mission-control — a sci-fi HUD over a TypeScript orchestrator that routes your intent to skills, runs them on the model of your choice (cloud or local), and writes every result to an Obsidian vault as the durable source of truth. Inspired by the "Jarvis" pattern: talk (or type) to it, it does the work in the background, and tells you what it found.

Status: active development. The gateway, routing/model layer, skills, memory vault, and HUD are working end-to-end. The Python voice sidecar is a scaffold (text mode is the default and works everywhere).

• Two-path routing — a deterministic regex table for known commands (instant, no LLM), falling back to a semantic brain (Haiku by default) for free-form intent. The brain is fast/cheap and orchestrates routing only.
• Pluggable models, per skill — each skill declares a modelPolicy; the selector picks a provider (Anthropic via API or headless claude -p, Ollama, or any OpenAI-compatible endpoint) by tier, privacy, budget, reachability, and a configurable fallback order. Headless Claude needs no API key, so it runs out of the box.
• Skills — declarative manifests (deep research, inbox triage, ship-ticket…) that run as native TS handlers, headless claude -p sessions, local processes, or composites. Results are written to the vault.
• Obsidian vault memory — every result is a clean, human-first markdown record with provenance frontmatter, a spoken-core TL;DR, cited sources, and graph backlinks. A freshness guard serves fresh records instantly instead of re-running.
• The HUD (V.A.U.L.T.) — an animated particle core, drag-arrangeable widget panels, a command deck built from your skills, task-completion notification cards orbiting the core, a markdown result viewer with "Open in Obsidian", an Audio I/O widget, and a live Options panel.
• Voice (opt-in) — text by default; flip to voice and the OS reads finished tasks aloud (spoken core). Local (faster-whisper/Kokoro) or cloud engines.

See docs/ for the full architecture and contracts.

Requires Node 20+ and Claude Code logged in (for the zero-config headless brain). npm install at the repo root wires the workspaces.

npm install

# Terminal 1 — the gateway (routing, skills, memory, HUD feed) on :7777
npm run start

# Terminal 2 — the HUD dev server on :5173
npm run hud

Open http://localhost:5173. Type a command (e.g. "last 30 days on rust") or click a card in the command deck. Results stream into the V.A.U.L.T. feed and open in the viewer; the file also lands in vault/.

No API key needed: the default transport is headless (claude -p, your local Claude Code login). Add keys/providers later in the Options panel.

apps/hud/            Vite + React + TS — the V.A.U.L.T. HUD (WebSocket client)
services/gateway/    Node/TS orchestrator: routing · models · skills · memory · WS server
services/voice/      Python voice sidecar (faster-whisper + Kokoro) — scaffold
packages/shared/     Shared contracts (events, skill manifest, model policy, vault schema)
config/              Config (file → env override) + secret store (OS keychain / encrypted)
skills/              Declarative skill manifests (+ ship-ticket = the old Shape A flow)
vault/               Obsidian vault — generated records are git-ignored
docs/                Architecture & contract docs

It's an npm-workspaces monorepo (@aos/*). TypeScript strict, ESM.

(See package.json for the full demo set: runtime-, deck-, voice-, mail-, mail-auth-, config-demo.)

• Architecture — components, data flow, the event bus
• Configuration & secrets — precedence, keychain, providers
• Model & provider layer — brain vs skill execution, fallback
• Skills — manifest schema, surfaces, model policy, native handlers
• Vault & memory — the document contract, freshness, spoken core
• Voice & HUD — modes, Audio I/O, widgets, cards