Verdict
Nanobot on a Pi 5 (8 GB) is a cheap way to run a personal agent: it ships Linux ARM64 wheels, and you can pair it with Ollama running Phi-3 mini for an entirely local stack. Its bubblewrap sandbox and workspace guard are off by default, and it is built for one owner: switch the sandbox on and match the rest to your threat model first. Until October 2026 this page said Nanobot has no sandbox by design; it has one you switch on.
Setup notes
64-bit Pi OS Bookworm, Python 3.11, virtualenv. Install nanobot-ai from PyPI. Install Ollama for ARM (the Pi 5 builds work), pull Phi-3 mini 3.8B Q4 and point Nanobot's ollama provider at it. Install bubblewrap (apt install bubblewrap) and set tools.exec.sandbox to "bwrap".
Performance
No measurement of ours. Phi-3 mini is a 2.2 GB download (phi3:mini in the Ollama library), and the Slips project measured 3B models at about 5 tokens per second on a Pi 5. On the 8 GB board that leaves headroom for tools and the OS.
What breaks
- A shared agent with per-user accounts
- Anything requiring 7B+ local model capability
Sources
The agent's defaults and settings above were read in its own documentation on 2026-10-18.
- Nanobot README (v0.3.5): install, platform wheels
- Nanobot docs (v0.3.5): providers, including Ollama
- Nanobot docs (v0.3.5): configuration, security section
- Nanobot SECURITY.md (v0.3.5): exec sandbox
Want to know more
See the Nanobot page and the Raspberry Pi 5 spec sheet.