Knowlin

One mind, many bodies.
A self-learning companion that embodies small robots –
virtual in the simulator today, then physically in your real surroundings.

Quick start · Features · Bodies · Safety · Contributing · Deutsch

Just want a look? Pictures and an overview at knowlin.de.

The idea

Knowlin is one shared mind for several robots. It knows your surroundings, your everyday life and the things you have experienced together – and takes on a body depending on the task: the walking Microduck, the self-built Open Duck Mini, the desk robot Reachy Mini or the SO-101 gripper arm at its workbench. Each embodiment has its own name, voice and character; the knowledge is shared. You choose the language model freely – Claude, another model with your own key, or a local one on your PC.

Without a body, Knowlin is calm and matter-of-fact. When something needs movement – walking, dancing, showing you something – it suggests the right robot and asks whether to use it.

What Knowlin does

Voice control Free sentences, wake word or push-to-talk: the language model understands, plans sequences and rehearses new moves invisibly in the physics first – Claude, Gemini, OpenAI, Mistral, or local with Ollama and LM Studio. Simple commands also work without AI.	Memory and experience network Knowlin remembers places, things, tricks and what you tell it – and where it learned it: without a body or in which embodiment, in the simulator or on the real robot. Everything stays on your PC.
One character, many variations Pick Knowlin's basic character on a field between calm and lively, serious and playful. Each embodiment has its own variation and the traits of its form – from beak clattering to antenna play.	Dancing, games, everyday life Dances on the beat, Pollen's emotions, camera games with Reachy (“Red light, green light”), timers, reminders and appointments – and a phone remote on your own Wi-Fi.

_{Images with Microduck and Reachy Mini: Knowlin with Pollen Robotics' 3D models – © Pollen Robotics, CC BY-NC-SA 4.0. These images are under the same licence; Knowlin does not ship the model files themselves.}

Also: its own voice running locally (Piper, or a sound designed with Qwen3-TTS) · speech recognition and chatting · simulator with a living room, things to play with and optional real physics · tricks from the Hugging Face Hub, courses and training (free on your PC or on Kaggle) · Knowlets: share character, tricks and dances as a package – data only, never knowledge about people · interface in English and German.

A mind, not just a language model

Knowlin stays the same, whichever model is thinking and whichever body is attached:

• Values and limits: every language model gets the same values (be honest, safety before curiosity, don't manipulate …). The limits – emergency stop, lock, no moving around on its own on the real robot – live in the code, each backed by a test. See Character → Personality.
• Intention, not movement: the model says what it means ("greet", "be happy"); the body works out how to show it – the duck nods, Reachy uses head and antennas, the arm waves. What it can't do, Knowlin says honestly.
• Self-image: "How are you?" is answered from the body's data, not made up.
• Drive and trail of thought: curiosity, closeness, rest and skill help decide what it does on its own – subtle, never pushy. Every action it takes on its own is listed with a reason.
• Looking back: in quiet phases it notes small lessons ("likes practising tricks"). It doesn't bring them up; you can correct or delete each one.

Bodies and modules

The modules have code names by kind of body: duckpilot for the walking ducks, deskpilot for stationary desk robots. Every body shows only what it can do – what it lacks (legs, depth sensor, beak …) simply isn't there.

Your own robot: With the body interface any robot can get Knowlin – as its own Python package, without changing Knowlin. It describes what it can do in a profile and passes the safety checks (python -m duckpilot.koerper_pruefung). Template: examples/knowlin-koerper-vorlage.

flowchart LR
  Human(["You – voice, text, phone"]) --> Mind
  subgraph Mind["Knowlin – one mind"]
    direction TB
    Brain["Brain<br/>Claude · other models · local · simple commands"]
    Values["Values, self-image, drive, looking back"]
    Memory["Memory and experience network<br/>surroundings · everyday life · experiences"]
    Character["Basic character and voice"]
  end
  Mind --> Safety{{"Safety<br/>emergency stop always wins"}}
  Safety --> D["duckpilot<br/>Microduck · Open Duck"]
  Safety --> R["deskpilot<br/>Reachy Mini"]
  Safety --> A["Arm<br/>SO-101"]
  D --> Sim[("Simulator<br/>with physics")]
  R --> Sim
  A --> Sim
  D -. "switchable" .-> Real[("Real robot")]
  R -. "switchable" .-> Real

Quick start

Windows – with the installer (recommended)

1. Download Knowlin-Setup – under “Assets”, the file Knowlin-Setup-….exe.
2. Double-click it. Python is included, no administrator rights needed. If Windows says “Windows protected your PC”, click More info → Run anyway – the installer isn't signed yet.
3. Start Knowlin from Knowlin in the start menu or on the desktop. The First steps assistant walks you through language, body, brain and the safety rules. For Claude, paste your Anthropic API key (platform.claude.com → Settings → API keys). It is stored in the operating system's vault, never in a file. Without a key, simple commands work (“walk forward”, “dance”, “stop”).

Updates: just install the new installer on top – memory and settings stay. Uninstall via Windows Settings → Apps.

With Python (Windows, macOS, Linux): pipx install knowlin, then knowlin --app (own window) or knowlin (console and browser). On Windows, knowlin --verknuepfung adds the start menu and desktop entries. Update with pipx upgrade knowlin.

From source: clone the repository, then on Windows double-click start.bat. The first time it creates a Python environment under %LOCALAPPDATA%\duckpilot (not in your OneDrive folder) and installs the packages and shortcuts.

On macOS and Linux from source: ./start.sh (console and browser) or ./start.sh --app (own window with Chrome/Edge). Needs Python 3.11 or newer; on Debian/Ubuntu also python3-venv. The browser's speech recognition only exists in Chrome and Edge – on Linux, use local recognition instead (Settings → Conversation → Speech recognition “Local”, after a one-time .venv/bin/pip install faster-whisper).

Safety and privacy

Please report security issues privately: SECURITY.md.

Accessibility

The goal is WCAG 2.2 AA: fully usable by keyboard, contrast checked, “reduce motion” respected. Every view is checked automatically with axe-core; testing with real screen readers (NVDA, VoiceOver) is still to do – feedback is very welcome. Details (German): docs/barrierefreiheit.md.

Documentation

• User manual (German) – every area and feature in detail
• Body interface – how another robot gets Knowlin
• Code signing policy – how the Windows installer is built and (soon) signed
• ARCHITECTURE.md (German) – structure, layers and where things belong
• CHANGELOG.md (German) · Release notes (English and German) · Roadmap

Contributing

Contributions are welcome – code, translations, dances, tests or reports from a real robot. Start with CONTRIBUTING.md and the issues labelled good first issue. Questions and show & tell are welcome in Discussions. Please follow the Code of Conduct.

Legal and sources

Knowlin is a free, non-commercial hobby project for small robots. It is not affiliated with, endorsed or reviewed by Pollen Robotics or Hugging Face. Knowlin is licensed under the Apache License 2.0 (LICENSE), © 2026 The Knowlin Authors; built with Claude (Anthropic). “Microduck” and “Reachy Mini” are names of Pollen Robotics; “Open Duck Mini” is a project by Antoine Pirrone. Contributions and their licences are listed in NOTICE and in the app under System → Settings → About. On GitHub: knowlin-ai/knowlin – maintainer @knowlin-dev.

All changes per version: CHANGELOG.md (German).