LYRA

A local-first personal AI operator you can talk to. Crimson void HUD, volume R3F orb, Daily / Complex modes, Orchestra / Obsidian Brain, Gemini-or-any OpenAI-compatible model, and permissioned tools on your machine.

Identity lives in your notes. After clone, point LYRA at your Brain and your keys.

Prototype 1 is a working operator on this kernel. The glass is a harness. Frontend heaven is a later Fable pass — do not treat the current chrome as the final look.

What works today

Honest list.

• Volume nucleus in the original LYRA orb (R3F). Screen 1 crimson HUD harness
• One session for type and talk (LyraProvider)
• Daily vs Complex tool budgets. Daily: Brain, web, fetch, calc. Complex: disk, GitHub, writes (HITL)
• DEMO replies until a model key exists and the MODEL chip is LIVE. No fake READY. Live demo: false is proven against a local OpenAI-compat /v1 (OmniRoute). Clones still bring their own key.
• Push-to-talk (hold Space / HOLD TO TALK), Web Speech STT/TTS, Esc barge-in. Interim transcript under MIC OPEN. Voice chip is VOICE · BROWSER · FEMALE after a browser probe — no TTS HTTP key required
• After tools: orb VERIFYING then brief SUCCESS, then STANDBY
• Brain search with file provenance. /remember and /capture propose writes (inbox.md, capture/*.md, projects/*/progress.md). Confirm on the desk. LYRA_AUTO_WRITE stays false
• The model can call tools. Slash commands are shortcuts, not the intelligence
• C: disk list/read with a secret denylist (ssh, .env*, .aws, Windows, browser profiles, private keys). Writes wait for confirm. Not a shell
• GitHub write is preview-only. MCP HTTP adapter if LYRA_MCP_HTTP_URL is set
• Settings (/settings): connections without secrets. OS alarm · UNAVAILABLE on purpose
• Work dock (O): TASK / SOURCES / WATCH. Chat hide (H). YouTube embed only with a real URL
• Local ledger in .lyra/. GitHub Actions: npm ci, npm test, typecheck, build

Not claimed: always-on mic. Live GitHub mutations. Unrestricted shell. Windows alarms. Phone pairing. Spotify OAuth. Frontend heaven.

Run

git clone https://github.com/mahik504/lyra.git
cd lyra
npm install
copy .env.example .env.local
npm run dev

Open http://localhost:3000.

Until LYRA_LLM_BASE_URL and LYRA_LLM_API_KEY are set, chat stays DEMO · NOT ONLINE. That is success of honesty.

Model gateway

Point LYRA at any OpenAI-compatible /v1. Local OmniRoute is the path that passed Prototype 1 live tests on this machine (Gemini Studio keys were unreliable here). Never commit .env.local. Never paste the key into chat.

LYRA_LLM_BASE_URL=http://localhost:20128/v1
LYRA_LLM_API_KEY=your_omniroute_dashboard_key
LYRA_MODEL_FAST=auto/best-free
LYRA_MODEL_BALANCED=auto/llama
LYRA_MODEL_REASONING=auto/best-coding
LYRA_MODEL_CODING=auto/best-coding

Daily uses LYRA_MODEL_FAST then BALANCED. Complex uses LYRA_MODEL_REASONING then CODING then BALANCED.

Direct Gemini (Google AI Studio) still works as an alternative:

LYRA_LLM_BASE_URL=https://generativelanguage.googleapis.com/v1beta/openai/
LYRA_LLM_API_KEY=your_studio_key
LYRA_MODEL_BALANCED=gemini-2.5-flash

Use a Gemini model name, not gpt-4o. Docs: OpenAI compatibility. Groq (https://api.groq.com/openai/v1) is another fallback.

Voice READY is Chrome/Edge Web Speech. Optional cloud TTS later via LYRA_TTS_*.

Controls

• Drag orbit, scroll zoom, R reset, O work dock, H hide chat, / focuses the field
• Hold Space (when not typing) or HOLD TO TALK. Esc / STOP barge-in
• Daily / Complex toggle on the desk
• Shortcuts: /search /fetch /github /work path /play URL /remember note /capture /gh-write — or just ask; the live model picks tools

Tests

npm test
npm run typecheck
npx playwright test

Playwright is local (needs a browser). CI does not run it.

License

MIT. Copyright Mahaveer Singh Gehlot. See LICENSE and NOTICE.md.