REMY — Assistente de Voz que Reescreve Seu Próprio Código
AbhigyaGoel/jetsonclaw
Assistente de voz local em Jetson que evolui escrevendo suas próprias skills
Ver no GitHubUm assistente de voz local em Jetson que executa wake word, speech-to-text e chat totalmente no dispositivo, enquanto delega tarefas complexas a sessões Claude Code headless. A skill revolucionária: reescreve seu próprio código sob demanda, com testes automáticos e auto-revert em loops de crash.
Instalação
npx convez-ai AbhigyaGoel/jetsonclaw1 comando no terminal: detecta o formato do repositório (marketplace ou plugin avulso) e instala via Claude Code. Requer Node 18+ e o Claude Code instalado.
Métricas
Casos de uso
- Automação de casa inteligente com evolução contínua de capacidades
- Assistente pessoal em dispositivo que aprende e cria novos comandos
- Desenvolvimento iterativo de skills sem reimplementação manual
Veredito

Projeto ambicioso e bem arquitetado: duas camadas de cérebro (local + agent remoto) eliminam latência em comandos triviais enquanto mantêm poder computacional real. O mecanismo de auto-modificação com testes e revert é elegante e seguro; ideal para quem quer um assistente que cresce sem intervenção.
README
A tradução para o português deste README ainda está na fila.
REMY
REMY (Recursively Evolving Machine, Yours) is a voice assistant that runs on a Jetson Orin Nano and can rewrite its own code. Wake word, speech to text, and chat run fully on device. Anything that needs real work spawns a headless Claude Code session billed to your existing Claude subscription, not an API key.
» "give yourself a skill to flip a coin"
« "That'll modify my own code. Say yes to proceed."
» "yes"
« "On it. Give me a few minutes."
agent: Creating a coin flip skill at ~/.remy/skills/coinflip/
« "Done. Say 'flip a coin' and I'll respond with Heads or Tails."
» "flip a coin"
« "HEADS!"
Real session, unedited. The skill existed before the sentence finished.
How it works
mic > openWakeWord > faster-whisper > intent router > skills (Spotify, self-grown) local, <1s
> qwen2.5 via ollama local chat
> Claude Code headless agent tasks
EventBus > TUI / web dashboard / logs
responses > Piper TTS (streamed sentence by sentence)
Two brains on purpose. Quick commands never leave the room. Code changes and integrations go to a real agent.
Quick start
# on the Jetson (or any arm64 Linux with a mic)
git clone https://github.com/AbhigyaGoel/remy ~/remy
cd ~/remy && bash scripts/install.sh
# local chat model
curl -fsSL https://ollama.com/install.sh | sh && ollama pull qwen2.5:3b
# agent brain auth: run on any machine with a browser, paste token to the Jetson
claude setup-token
echo 'CLAUDE_CODE_OAUTH_TOKEN=<token>' >> ~/.remy/env
# check everything, then run
~/.remy/venv/bin/python -m remy --doctor
~/.remy/venv/bin/python -m remy
Dashboard at http://<jetson-ip>:8484. Install it as a PWA on your phone: it has a mic button, so your phone is a remote.
What it does
| Say | Result |
|---|---|
| "play blinding lights" | Spotify search and play |
| "play my gym playlist" | fuzzy playlist match |
| "what time is it" | a self-grown skill answers |
| "remember that my locker code is 4912" | written to long-term memory |
| "what do you remember about my locker" | recalled from memory |
| "give yourself a weather skill" | writes, tests, and hot-loads a new skill |
| "upgrade yourself to ..." | edits its own repo, gated by tests |
| "undo that" | reverts the last self-change |
| "status report" | uptime, skills, memory, RAM, disk |
| anything else | local chat with conversation memory |
Typed input works everywhere (TUI, dashboard, stdin) and runs the same pipeline as voice.
Self-modification, safely
- Every change runs in a fresh Claude Code session inside the repo
- The change must pass
python3 -m remy --selftestor it is discarded - Passing changes are committed; the previous commit is recorded as last known good
- A crash loop at boot auto-reverts to last known good after 3 failed starts
- The agent has no shell access by default and asks for a spoken yes before running
Skills are simpler: a SKILL.md plus optional handler.py under ~/.remy/skills/, hot-loaded on the next utterance with no restart. Declared pip dependencies are installed by the harness, and a failing selftest quarantines the skill. See docs/skills.md.
Memory
One episodic store, three views: the last few turns feed chat context, keyword search recalls older interactions, and an idle-time consolidation pass summarizes each day and folds durable facts into MEMORY.md. Personality lives in ~/.remy/SOUL.md and is plain markdown the agent itself can edit.
Hardware
- NVIDIA Jetson Orin Nano (8GB), JetPack r36, Ubuntu 22.04
- any USB mic (tested with a Logitech C270 webcam)
- any speaker (HDMI or USB)
Other arm64 or x86 Linux boards should work; the Jetson is what it is tuned and tested on.
Providers
The chat brain speaks two protocols: ollama (the on-device default) and OpenAI chat completions. One config block covers OpenAI, Groq, OpenRouter, Together, DeepSeek, Mistral, vLLM, llama.cpp, LM Studio, and anything else with a /v1/chat/completions endpoint:
[chat]
provider = "openai"
url = "https://api.groq.com/openai/v1/chat/completions"
model = "llama-3.3-70b-versatile"
api_key_env = "GROQ_API_KEY"
Configuration
Copy config.example.toml to ~/.remy/config.toml. Assistant name, wake word model, voices, models, ports, and agent permissions are all config.
Docs
| Doc | Covers |
|---|---|
| docs/features.md | everything REMY does, and how to remove what you don't want |
| docs/skills.md | skill format, synthesis, activation, quarantine |
| docs/wake-word.md | training a custom wake word |
| docs/jetson.md | Jetson-specific setup notes and pitfalls |
Cost
Wake word, STT, TTS, chat, Spotify, and skills run locally and cost nothing. Agent tasks draw from the Agent SDK credit included in Claude subscriptions (about $100/month equivalent on Max 5x as of mid 2026).
License
MIT