Homebody for Reachy Mini
AI-generated promotional illustration—not a product photograph. Image credits.
Homebody gives Reachy Mini a home: a voice companion for everyday questions, explicitly selected project work, and supervised games and stories. The local app owns the body, dashboard, power and privacy controls; an optional private bridge connects speech providers and scoped host capabilities.
Explore the app page ↗ · Hardware · Setup · Models and speech routing · Security · Contribute
Early alpha. Reference deployment is not broad hardware or production acceptance. The original Lite + Raspberry Pi 4 reference is retained; the current Lite + Jetson host has app, camera, Home Assistant, power, speaker and managed-update reference deployment. Each installation still needs supervised wake/fold, camera/privacy, speech and acoustic interruption checks. Local GPU vision, accelerated gestures, GPIO and controller support are not blanket verified on Jetson.
Release boundary: published releases, source and installed robot/bridge versions are distinct; see release status and the installed identity. Version 0.5.15 adds the first marketplace-aware onboarding slice, not complete terminal-free installation or broad platform acceptance. Packaging companion source does not configure or restart a subscription bridge or native Workspace.
Choose the purpose, then the speech engine
| Purpose mode | What it is for | Boundary |
|---|
| Everyday | General explanations, configured weather/time and allowlisted home controls. | Narrow host backend: no full agent, private memory, project access or full-agent fallback. |
| Workspace | Discuss a selected project and, after approval, run registered project work. | Native Hermes requires explicit host project/session registration and owner selection. Broker-only discussion is labelled separately. |
| Kids | Supervised buddy chat, stories, quizzes, riddles, calm-down time and I Spy. | Dedicated route with no adult memory or tools. Moderation, per-session I Spy camera consent and Stop/expiry/fold lifecycle remain in force. |
These are not speech engines. Under Advanced settings, Pipeline uses STT → text → TTS; Realtime uses OpenAI speech-to-speech; GPT-Live preview is an experimental speech relay. GPT-Live exposes no direct camera, movement or memory tools; delegated text follows the selected purpose boundary; physical echo cancellation and barge-in acceptance remain open. See Everyday, Workspace and voice engines and Kids Mode.
Native Workspace preserves the selected Hermes session's identity and history. The planned execution flow needs exact phone approval, registered verification commands and artifact receipts; a working directory is context, not a sandbox. Native discussion uses the worker's configured tools, so “discuss only” is an instruction, not enforcement, and planned-execution approvals do not universally gate discussion tool calls. Configure owner-scoped read-only tools when that guarantee is needed. Stop speaking is distinct from Stop project work; returning to Everyday can park approved work while hiding its private display. See native Workspace.
Hermes Agent is the native reference backend. OpenClaw conversation compatibility remains available, but is not native Workspace or subscription feature parity.
Companion: notice, help, then go quiet
The Everyday Companion card brings presence, deliberate Look at this, Stop and opt-in suggestions together. One explicit look captures one frame on the selected local-vision or host route; bounded follow-ups reuse its text without another picture or household/project tools. Stop, privacy, permission changes and owner revocation discard pending results. Temporary Everyday camera grants remain bounded and revocable.
Suggestions stay silent until you choose a source and enable the relevant policy. Source enrollment uses only host-authorized identifiers; calendar details are private on the owner dashboard by default, not read aloud to an anonymous “yes”. Occupancy is not identity or attention, and no source enrollment grants device control. Existing settings and Kids protections remain in force. See Companion setup and boundaries.
What can I expect?
Start with the local robot, then add connected capabilities deliberately.
- Set up Reachy Mini Wireless, or a wall-powered Lite with USB data. For community hosts, read the hardware guide and Lite + Pi 4 guide.
- In Reachy Mini Control, choose Applications → Discover Apps → Homebody → Install → Start. Use Reachy SDK 1.9.0+ and Python 3.11+; managed updates have additional interpreter requirements in the update guide. See the marketplace quick start and first-run limits.
- Open
http://REACHY_HOST:8042 on your trusted management network. Pair an owner browser using the documented pairing flow; use private trusted HTTPS for the phone PWA.
- On a fresh 0.5.15 install, choose Continue setup: confirm the host, configure the backend, check its connection silently and review next-Start listening. The app stays quiet until setup is saved and you explicitly Stop/Start it. Initial secure owner provisioning and standalone backend installation remain separate setup work. Existing saved installations retain their startup behavior. Follow operations and acceptance for supervised speech, guarded movement, camera/privacy and fold/torque checks one at a time.
The local dashboard works without an agent. Meeting and Sleep stop app microphone capture and wake detection. Camera viewing is opt-in and starts only after Start; the private HTTPS frame fallback has a lower frame rate than WebRTC. Newer camera lifecycle/startup repairs and safe movement controls do not replace supervised acceptance. See robot controls.
Supported setups
Wireless, Lite + Linux/Windows PC, Lite + Mac, and community Lite + Pi/Jetson hosts have different capability and acceptance boundaries. The hardware matrix separates original Pi evidence, deployed Jetson core features and experimental accelerators/controllers; no whole-platform green check is implied.
Install the Reachy app
For development from a local checkout:
uv sync --group dev
uv pip install -e .
Start through the Reachy dashboard. For managed appliances use the managed-update procedure, not an ad-hoc wheel swap. It updates Reachy software and dashboard assets, not an Android APK. Owner pairing must use the running app's user and effective HOME/owner database; see browser pairing.
Connect speech and host capabilities
Install your own household MCP server (unreleased preview, no Hermes required for Everyday):
start with the installation runbook, including copyable
AI-agent instructions and a read-only smoke check. This is a separate development trial,
not a drop-in replacement for the shared Kids/Pipeline/Workspace bridge. See the
protocol and current limits.
Basic connection: Settings → Connect your agent creates a one-time message for Hermes Agent or OpenClaw. Review the plan before approving it. The agent downloads the robot-served bridge, checks its manifest, creates a restricted conversation profile and pairs. Agent-led setup explains the security model and manual alternative.
This is not zero-setup subscription or native Workspace provisioning. The bundle follows your installed robot release. It does not automatically configure saved Codex authentication, install subscription entry points or register native project sessions. Use the explicit subscription and native setup guide, alongside the companion bridge guide.
Optional subscription-funded text/vision uses saved Hermes Codex authentication on the trusted host, without paid text API fallback. Models must genuinely accept reasoning none; low/minimal is not off. The validated reference configuration uses gpt-6-sol for Everyday, planning and native Workspace, and gpt-6-luna for Kids/I Spy, with reasoning none. These are reference choices, not universal defaults; gpt-6.1-sol currently does not support none.
Speech is separate: Realtime/GPT-Live and configured STT/TTS can incur API costs. Kids moderation remains an API safety-screening path; a subscription does not remove it or make the whole robot API-free. Provider credentials stay on the host, not the robot or dashboard.
Local models on your PC or Mac: Ollama setup covers Everyday Pipeline reasoning and camera vision, either on the same computer or over your trusted LAN. Ollama does not replace speech providers or Kids moderation, and the computer must remain awake.
Practical guides
Home Assistant integration is opt-in, with separate controls, camera and Assist permissions. Keep its ESPHome listener on a trusted LAN and firewall it to the HA host; do not expose the plaintext device API publicly. Capability allowlists start empty. Read security before sharing endpoints beyond a trusted LAN/VPN.
Architecture
Tap or open the diagram for its full-size labels.
The robot's lifecycle and camera permissions remain local. Everyday does not inherit Workspace authority; native Workspace is an explicit host opt-in. Camera descriptions can use an optional trusted local vision server, but local inference and accelerator performance need acceptance on your hardware.
The official Lite assembly preview provides hardware context, not a Homebody acceptance result.
Official Pollen Robotics image, converted to metadata-free WebP under Apache-2.0. Images are explanatory, not evidence of acceptance on every setup. Credits and immutable sources.
Development and credits
uv run ruff check .
uv run pytest
uv build --wheel
reachy-mini-app-assistant check .
See contributing, changes, operations and security. Homebody was previously Reachy Mini Hermes; upgrade notes preserve existing configuration.
Built by Tim Verhoogt, with assistance from Hermes Agent by Nous Research, on the work of Pollen Robotics and the Reachy Mini community. Third-party notices and model notices identify upstream code and models.
Apache License 2.0. See LICENSE.
