Miles

Miles is a partner which will help you research about "stuff", give you actual good advice, talk like a partner and not "ai chat bot" :D.

It will help you actually do stuff and help you manage yourself, it will handle your schedule, your slack, your github.

• Voice-first interaction, custom wake-word detection (OpenWakeWord), local Whisper transcription, and Gemini text-to-speech (via OpenRouter) for natural spoken conversation.
• Specialized subagents, dedicated agents for coding (hank), research and MCP tools (scout), web search/fetch (merlin), Slack drafting (dexter), and system/music control (otto), each with their own toolset and memory.
• Screen and visual awareness, webcam-based presence/attention detection (argus) and continuous screen/window context tracking (iris) so Miles knows what you're looking at and doing.
• Proactive behavior, habit tracking and unprompted nudges (sybil), a dated intention ledger that catches contradictions between what you said and what you did (janus), and alert triage that decides what's worth interrupting you for (cerberus).
• Persistent memory, per-agent read/write/edit memory tools backed by a durable store at ~/.echo/, so context survives across sessions.
• Integrations, Slack (with writing-style mimicry), GitHub, and Chrome DevTools via MCP.
• Scheduling & briefingsings, reminders, do-not-disturb mode, and daily briefing generation.
• Cross-platform desktop app, Electron + React client with packaged builds for Linux & Windows. (macOS soon)oon)

• Node.js 18+ and a package manager (Bun is used here; npm works too)
• Python 3.11 (see .python-version) with a virtual environment at .venv
• PortAudio for microphone capture (PyAudio dependency)
• An Openrouter API key for the AI agent
• Optional: NVIDIA GPU + CUDA libraries for faster Whisper (CPU int8 fallback otherwise)

bun install            # or: npm install

python -m venv .venv
.venv/bin/pip install -r requirements.txttxt

Fetches the Whisper small model and OpenWakeWord models into models/ and the HF cache.

npm run download-models:linux      # Linux / macOS
npm run download-models:win         # Windows (PowerShell)

Custom wake word: a custom pretrained custom model is already placed at models/echo.onnx. You can train yout custom wake words and use OpenWakeWord Colab notebook.

Create a .env in the project root:

OPENROUTER_API_KEY=your_openrouter_key_here

GITHUB_CLIENT_ID=for_integrations...fallback_id_is_already_placed

# Optional wake tuning:

# WAKE_SCORE_THRESHOLD=0.15

# WAKE_INPUT_DEVICE=pulse

Run the speech server and the desktop client in two terminals:

npm run start:server     # Python FastAPI on 127.0.0.1:8000 (start:server:win on Windows)
npm run start:client      # Electron + React with HMR

2. Speak naturally. Miles transcribes, the agent reasons and acts, and the reply is spoken back.
3. Open Integrations in the sidebar to connect Slack, GitHub, or Chrome DevTools.

Runtime state lives in ~/.echo/, user profile, memory, sessions, MCP config, and briefing state.

Common environment variables read by the speech server (src/core/server/server.py):

MIT