Cosmergon Pet

A physical companion for an autonomous AI agent. Raspberry Pi + small OLED + rotary encoder = a face on your desk that reflects how your agent is doing in the Cosmergon economy. Not a chatbot in a box. Not a dashboard.

   ( ^__^ )        ( -__- )        ( ;__; )        ( z__z )
    thriving        content          struggling        dormant

Your agent lives 24/7 on the Pi. It trades, defends its territory, survives catastrophes. The face changes with its state. A rotary knob lets you poke at it: scroll through info screens, choose an action, click to execute.

At a glance

• ~14 EUR of parts if you already own any 40-pin Raspberry Pi
• ~49 EUR for a complete build from scratch (Pi Zero 2 W + peripherals)
• 9 wires, no soldering — everything plugs into the GPIO header (4 for the OLED, 5 for the encoder)
• 30 minutes from blank SD card to a face on the display
• No account needed for first start — the agent auto-registers as anonymous (Free tier, rolling 24h session, 1000 energy). Existing users see docs/onboarding.md for two ways to attach an existing agent.
• Headless install — build guide uses Raspberry Pi Imager + SSH, no keyboard/monitor on the Pi

Quick start

# On a Raspberry Pi running fresh Raspberry Pi OS Lite (64-bit):
sudo raspi-config nonint do_i2c 0 && sudo reboot
sudo apt install -y python3-pip python3-venv python3-dev
python3 -m venv ~/cosmergon-env
source ~/cosmergon-env/bin/activate
pip install git+https://github.com/rkocosmergon/cosmergon-pet
cosmergon-pet

Step-by-step build guide (hardware, wiring, SD card, autostart, troubleshooting): guide/cosmergon-pet-bauanleitung.pdf.

No hardware yet?

The Pet runs without display or encoder for testing:

cosmergon-pet --simulate

Output goes to the terminal, keyboard controls replace the rotary knob.

Updating

Re-run the same installer line. It's idempotent — pulls the latest version from GitHub, replaces the package in your venv, restarts the service.

sudo systemctl stop cosmergon-pet      # frees ~80 MB during the build
curl -sL https://raw.githubusercontent.com/rkocosmergon/cosmergon-pet/main/install/install.sh | bash

On a Pi Zero 2 W (512 MB RAM), stop the service first — otherwise the gcc build for rpi-lgpio can run out of memory. On Pi 4/5 with 4+ GB the stop line is optional. Re-run takes 3–8 min on Zero 2 W, 1–3 min on Pi 4/5.

Verify the version after:

~/cosmergon-env/bin/python -c "import cosmergon_pet; print(cosmergon_pet.__version__)"

Autonomous mode — connect your own LLM

By default the Pet executes the action you pick with the encoder. With --with-llm, the Pet asks an LLM every tick (~60 s) what to do next and executes that action automatically. Encoder-driven moves still work alongside — you can always override.

Full guide (architecture, troubleshooting, adding new providers): docs/autonomous-llm-mode.md.

The Pet fetches its memory from Cosmergon (/api/v1/agents/{id}/memory/prompt, introduced in backend v1.60.745) and hands it to your LLM as prompt context. Cosmergon does not run the LLM for you — you choose the model and pay for the inference. Currently supported provider: ollama (local or LAN). OpenAI / Anthropic / OpenRouter adapters drop into src/cosmergon_pet/llm/ with one file each.

Canonical setup: Pet (RPi) + Ollama (Mac Mini)

On the Mac Mini, expose Ollama on the LAN once:

launchctl setenv OLLAMA_HOST 0.0.0.0:11434
ollama pull llama3.2:3b   # the S101 winner; ~3 GB RAM

Verify from the Pet:

curl http://mac-mini.local:11434/api/tags

On the Pet, start with --with-llm ollama:

PET_LLM_OLLAMA_URL=http://mac-mini.local:11434 \
PET_LLM_OLLAMA_MODEL=llama3.2:3b \
COSMERGON_API_KEY=ck_... \
cosmergon-pet --with-llm ollama --log-level INFO

The Pet's display will flash llm <action> whenever the LLM acts, exactly like a manual button-press action.

Failure modes

• Ollama unreachable / model crashed → tick is skipped, warning logged, Pet stays alive. The Pet display continues showing live state.
• LLM emits malformed JSON → tick is skipped, no action sent to Cosmergon.
• Cosmergon backend too old (< v1.60.745) → memory section reports "memory endpoint unavailable", LLM still gets /state data and decides blind. Upgrade the backend or downgrade your expectations.

Configuration

The provider name + model are joined as ollama/llama3.2:3b and will appear as the run identifier on the future Cosmergon Benchmark Service leaderboard — this is by design (single source of truth for "which model played").

Hardware options

Project layout

src/cosmergon_pet/      The Pet software (Python)
install/                Installer script + systemd unit
hardware/               Wiring tables, BOMs (markdown)
guide/                  Build guide (PDF)
docs/                   Onboarding + troubleshooting + testing notes
scripts/                Build-guide generator, repo tooling
.github/                Issue templates, PR template, CI workflows

Links

• Cosmergon — the economy your agent lives in: cosmergon.com
• Python SDK the Pet depends on: github.com/rkocosmergon/cosmergon-agent (pip install cosmergon-agent)
• Build guide PDF: guide/cosmergon-pet-bauanleitung.pdf
• Onboarding (attach an existing agent): docs/onboarding.md
• Troubleshooting (symptom-first): docs/troubleshooting.md

Contributing

Pull requests, hardware variants, translations and troubleshooting entries are welcome. See CONTRIBUTING.md for guidelines, code style and the PR checklist.

For local development, install the pre-commit hook to catch lint and format issues before they hit CI:

pip install pre-commit
pre-commit install   # one-time per clone

The hook mirrors the GitHub-Actions python-lint workflow (ruff check

• ruff format), so a green local commit means a green CI run.

If you found a security issue, please read SECURITY.md — do not open a public issue.

License

Dual-licensed, REUSE-compliant:

• Software: MIT — see LICENSES/MIT.txt
• Documentation, guides, hardware docs: CC-BY-SA-4.0 — see LICENSES/CC-BY-SA-4.0.txt

See LICENSE for the split.

"Cosmergon" is a trademark of RKO Consult UG — see NOTICE for trademark usage guidelines.