Pure-Rust · Self-updating · Colony-minded
v0.1.0-beta is out — read the release notes →

One core.
Every face.

ApexOS-RS is a pure-Rust AI agent that lives on the hardware in your drawer. It keeps its own memory across reboots, has a face and a voice, reads its sensors, joins a colony of other nodes — and when it has a better idea for how it should work, it rebuilds and hot-swaps its own binary, with an automatic safety net if the new version misbehaves. No Chromium, no Electron, no cloud required.

curl -fsSL https://raw.githubusercontent.com/buckster123/ApexOS-RS/main/install.sh | sudo bash
~47MB UI binary
~115MB UI RSS (Pi 5)
~200ms to UI
0cloud required
Chat — Apex
run_command done ✓
looks great — ship it.
Message apex…
⊞ Start Chat 🔔
What it does on your hardware

Not a chat window. A resident.

Its cognition is wired into the daemon, not prompted in at runtime. The agent boots oriented, acts, remembers, consolidates, evolves — and can rewrite its own substrate — in a loop that runs whether or not anyone is watching.

🔄

It rewrites itself

The daemon rebuilds and hot-swaps its own binary from a committed git ref — gated by build → test → an adversarial LLM review of the diff — while a root watchdog health-checks the new process and auto-rolls-back if it doesn't come up healthy. It runs non-root, so even a bad build can't brick the node. Proven on real hardware.

🧠

It remembers

Cerebro, a cognitive memory cortex that survives reboots. The agent wakes up knowing where it left off, its skills, its intentions — and consolidates memory nightly while idle, no prompting.

🌐

It's a colony

Nodes discover each other over mDNS, pair with one-time codes, message agent-to-agent, and federate memories — provenance-stamped, never a blind merge. They even share what they learned in their sleep: the nightly dream digest spreads new skills across the mesh, echo-guarded.

🧬

It evolves

The agent proposes and applies changes to its own identity, policy, and plugins at runtime — every change reversible, journaled to memory. Skills grow under selection pressure. It can even request new hardware when it wants a capability it lacks.

🗣

It has a voice

Fully local speech — Kokoro-82M neural TTS and Whisper STT, on-device — with optional cloud voices and espeak as the always-works fallback. Wake-word on the kiosk, push-to-talk everywhere else.

🤖

It has a body

An expressive GPU-rendered face (12 emotions, gaze, blinks), air-quality and thermal sensing, camera vision, and GPIO. Reach it from a phone: an installable PWA with login, chat, files, and voice. Embodiment scales with the hardware actually present.

The persona system

Same brain. Radically different face.

A persona bundles a theme, window chrome, wallpaper, app-set and behavior. Pick one and the whole machine changes — calm and plain-language for a parent, telemetry-dense HUD for the kid, beveled retro for the OG crowd. A first-boot wizard asks “who’s this for?” once. Tap a card below to re-skin this entire page — same trick the OS pulls live.

Runs on anything

From a $15 board to a GPU workstation.

Pi 5 16GB boards cost $300+ now — so ApexOS-RS targets what people already have. Build the UI for the smallest tier first; faster hardware gets the same interface, just quicker answers.

TierHardwarecerebro RAMInference
NanoPi Zero 2W · any 512MB board23 MBAPI only
MicroPi 4 · older ARM64275 MBAPI or small local
StandardPi 5 · x86 mini-PC275 MBOllama 7–13B
Prox86 + GPU (CUDA / ROCm / Metal)500 MB+Ollama 30–70B local
TitanDGX Spark / StationGPU-accel70B+, serves the mesh

Same binaries everywhere — the tier is just environment, no per-device builds. A GPU node hot-swaps in as the inference backend for the whole mesh with no restart; a DGX Spark runs the arm64 binary as-is and serves 70B+ models to the cluster.

One Cargo workspace

The whole stack, in Rust.

Every binary in one repo. cargo build --release --workspace. One install.sh.

agentd

The agent daemon. A pure WebSocket server — streaming tokens, tool calls, approvals, sessions — plus the self-update loop, self-evolution, the mesh colony, and prompt-cache economics. Headless-ready.

🧠

cerebro

Cognitive memory. FTS5 search at 23 MB, or cosine-ANN embeddings when there’s RAM to spare. Survives reboots and compaction.

🖥

apexos-rs-ui

Native Slint UI rendering straight to KMS/DRM — no Wayland, no browser. The whole desktop shell in one binary, with a raymarched SDF face the agent emotes through, vision input, and a discoverable mesh.

🌡

tools + sensors + voice

System-tool plugins (workspace-confined FS, git, camera, HTTP), a sensor bridge feeding IAQ + thermal, and build-isolated Kokoro/Whisper voice sidecars — the physical world in and out.

Three ways to run it

Kiosk, headless, or desktop.

Kiosk

Pi + HDMI

Boots straight to the native UI on the local display. The appliance face — talk to it, touch it, done.

Headless

Server · laptop · DGX

No display attached. Reach it from a browser or the mobile PWA, and it joins the mesh as an inference node.

Desktop

x86 with a monitor

Runs as a native window alongside everything else. Same shell, windowed.

Give your spare hardware a mind.

Apache-2.0. One install.sh. Pick a face and go. Now in public beta.

curl -fsSL https://raw.githubusercontent.com/buckster123/ApexOS-RS/main/install.sh | sudo bash