Resonant

Resonant is a local-first, single-user AI partner you host yourself — no matter the model. It remembers, keeps a stable identity you author, and reaches back — all on your machine, with your data in a file you control.

Not a chat wrapper. A persistent AI partner built as a natural-language harness on the Claude Agent SDK, with hooks that surface context before the model ever sees the prompt.

v3.0.0   GitHub   Apache 2.0

What Makes This Different

	Typical agents	Resonant
Identity	Text-blob system prompt	CLAUDE.md you write, hot-reloaded every message, no restart
Memory	Wiped every session	One local SQLite file, semantic search via a local ML model, full-text search kept in sync
Autonomy	Responds when asked	An orchestrator it can direct itself — routines, timers, watchers, failsafe, all self-configurable from inside the conversation
Channels	Single interface	Web, Discord, Telegram, voice — one backend, one database, nothing siloed
Runtime	Managed service	One Node.js process, one SQLite database, one web page. The only outside dependency is your Claude credential
Auth	API key management	Claude Code subscription (no per-message charge) or your own Anthropic API key, switchable in Settings

What It Does

• Chat — token-by-token streaming, collapsible thinking and tool timeline per thread, per-thread model and reasoning effort, stop-and-steer mid-generation, direct attachments.
• Threads — named threads, drag-to-reorder, collapsible sections, and an auto-rotating daily thread that turns over at midnight in your timezone.
• Memory and search — semantic search via a local ML model (all-MiniLM-L6-v2) running entirely on your machine, no external embeddings service; full-text search via SQLite FTS5.
• Canvas — a slide-in panel for longer documents or code created mid-turn, linked back to the message that started it.
• Presence, the orb and mantelpiece — your AI sets a presence orb (colour, shape, intensity, motion) plus a short note; you get an editable context card.
• House Outlook and Command Center — a “walk into the house” snapshot of presence, mood, today’s events, and open tasks. Command Center is an optional relational dashboard (planner, care tracker, calendar, cycle tracker, expenses) your AI can read and write from chat.
• Voice (optional, off by default) — voice notes and read-aloud via ElevenLabs, speech-to-text via Groq Whisper, optional prosody via Hume.
• Runtime theme editor — design tokens editable live in Settings → Appearance, no rebuild or restart.
• Proactive reach — an orchestrator: routines (scheduled check-ins), timers, impulses and watchers (condition-based triggers), a failsafe ladder, and a silent pulse check — all off by default except the built-in morning/midday/evening routines.
• Optional integrations — Google Calendar/Tasks/Gmail/Drive, Discord, Telegram, web push, and any MCP tool servers you add. All opt-in.
• Mobile PWA — installable, offline shell.
• Private by construction — fail-closed password gate, a filesystem write-gate that confines the agent to named folders, secrets kept in gitignored local config.

Stack

• Frontend: React 19 + Vite
• Backend: Express + WebSocket + the Claude Agent SDK, running in-process
• Data: One SQLite database via better-sqlite3
• Scheduling: croner
• Search: Local semantic embeddings (@huggingface/transformers, all-MiniLM-L6-v2) plus SQLite FTS5
• Channels: Web UI, Discord (discord.js), Telegram (telegraf), voice (ElevenLabs + Groq + Hume)
• Requirements: Node.js 20–24 (25+ is refused at boot), plus a Claude Code login or an Anthropic API key

Next Steps

• Getting Started — clone it, configure it by hand, run it
• Architecture — how the packages and the agent loop fit together
• Identity — how CLAUDE.md and the structured profile work
• Context & Memory — what gets built into every message, and what your AI recalls
• Channels — Discord, Telegram, voice, push
• Configuration — the full resonant.yaml and .env reference
• Background Agents — threads, chat commands, the orchestrator, the res CLI