Local AI that belongs to you. A tiny, dependency-free CPU inference engine and a governed agent built on it — private, offline, on hardware you already own, down to a Raspberry Pi. No GPU, no cloud, no Python at runtime.
Try it in one minute — Microsoft's BitNet 2B on a 4 GB Raspberry Pi 5, model included, offline after download:
curl -L -o geist-bitnet https://github.com/geisten/geistlib/releases/latest/download/geist-bitnet-linux-arm64
chmod +x geist-bitnet
./geist-bitnet "The capital of France is"geistlib is the engine; everything above it links it in-process at the
stable C ABI, never by patching the core. The engine has no plugin system
and no opinion about chat templates, tool use, or policy — each program on top
is small, self-contained, and owns its own product decisions.
geistshell (agent)
│
geistlib (engine)
| Project | Role | What it is |
|---|---|---|
| geistlib | engine | Dependency-free C23 inference engine for small LLMs — ternary BitNet, CPU-first, runs on a Raspberry Pi |
| geistshell | agent | Governed, auditable agent that learns from real runs — policy-gated and budget-bounded in C23; grammar-constrained decoding gets valid actions out of small models never trained for the format |
Install: brew tap geisten/tap — packages for macOS and Linux.