The smallest AI agent runtime for edge devices.
180KB
full: ≤500KB
Zero dependencies · No OS required · Any LLM provider
Runs on a $3 ESP32. Written in Zig.
zig build -Doptimize=ReleaseSmall export ANTHROPIC_API_KEY=... ./krillclaw "read the temperature and adjust the fan"🦐 I Want One!
What you get that others don't ship.
Choose the right build for your device.
💡 Lite and Full are independent builds. Use either one alone, or combine them in a fleet. Full does not require Lite.
| Feature | Lite (180 KB) | Full (≤500 KB) |
|---|---|---|
| Binary size | 180 KB | ≤500 KB |
| Transport | BLE / Serial | WiFi / HTTP / TLS |
| ReAct agent loop | ✅ | ✅ |
| Tool execution | ✅ local tools | ✅ full toolset |
| Persistent memory | ✅ flash | ✅ flash + filesystem |
| Config management | ✅ | ✅ |
| LLM API calls | via BLE gateway ⚡ | ✅ direct |
| TLS / HTTPS | not needed ⚡ | ✅ |
| GPIO / sensors | ✅ direct | ✅ |
| API key on device | ❌ secure! ⚡ | ✅ required |
| Power draw | ~10 mA | ~100 mA |
| Min hardware | $3 ESP32-C3 | $5 ESP32-S3 / Pi |
| Offline capable | ✅ local ops | ✅ queues calls |
⚡ = by design, not a limitation
BLE has 80× more bandwidth than an LLM stream needs.
The bottleneck is always the AI, never the transport.
How it comparesEmbedded AI agent runtimes — a new category.
| KrillClaw™ | MimiClaw | PicoClaw | OpenClaw | |
|---|---|---|---|---|
| Language | Zig | C | Go | Node.js |
| Binary size | 180 KB / ≤500 KB | ~2 MB | ~8 MB | ~200 MB |
| RAM usage | 2 MB | ~512 KB | ~10 MB | ~200 MB |
| Source | ~3,500 LOC | ~7K LOC | ? | ~40K LOC |
| Dependencies | 0 | ESP-IDF | Go stdlib | ~108 npm |
| Target hardware | $3 ESP32+ | $5 ESP32-S3 | $10 LicheeRV | Pi / VPS / Mac |
| Needs OS? | No | No | Linux | Linux / macOS |
| Memory safety | Yes (Zig) | No (C) | Yes (Go GC) | Yes (JS) |
| LLM providers | Any compatible | Claude + OpenAI | ? | Multi |
| BLE transport | Yes | No | No | No |
| Security | ✅ Built-in auth | ❌ None | ✅ Sandboxed | ✅ Full |
| Channels | 8+ | 1 | N/A | 8+ |
| Memory write | ✅ | ❌ | ✅ | ✅ |
| Media handling | ✅ Photos, voice, files | ❌ Dropped | ? | ✅ |
| Skills / plugins | ✅ ClawHub | ❌ None | ❌ | ✅ |
| CI / testing | ✅ 40 tests, 0 leaks | ❌ Zero | ? | ✅ |
| License | MIT | MIT | MIT | MIT |
You can read the entire codebase in an hour.
# Both profiles agent.zig — core loop (250 lines) json.zig — hand-rolled JSON (500 lines) context.zig — token management (225 lines) stream.zig — SSE parser (344 lines) transport.zig — BLE/serial abstraction # Full only api.zig — multi-provider HTTP (329 lines) tools_coding.zig — 7 tools (280 lines)
Works with any OpenAI-compatible LLM provider. Switch with a flag.
What becomes possibleWhen every device has an AI brain and WiFi, ordinary things become extraordinary.
A $3 ESP32 inside your fridge watches what gets consumed. It notices you're low on oat milk, checks your Instacart history, and pings you on WhatsApp: "You tried that new Greek yogurt last week but didn't reorder — want to go back to your usual Fage, or give it another shot?" It learns your preferences over time. No app. No screen. Just a chip that pays attention.
You're walking home with hands full of grocery bags. KrillClaw on your garage monitors the house vicinity via your security camera. It sees you approaching with your hands full, recognizes you, understands the situation, and opens the garage door as you walk up — so you don't have to awkwardly dig for the clicker.
Soil moisture drops. Temperature spikes. A traditional sensor triggers a fixed rule. A KrillClaw node reads the sensors, checks the 5-day forecast, recognizes the tomatoes are in flowering stage (needs less water, more heat), and adjusts the irrigation schedule accordingly. It messages you a weekly report: "Tomatoes on track for harvest in ~12 days. Reduced watering 15% — soil was retaining more than usual after Tuesday's rain."
A CNC machine's vibration pattern changes subtly. The KrillClaw node on the motor controller notices, cross-references the maintenance log stored in flash, and messages the floor supervisor: "Spindle vibration up 12% since Monday — similar to the pattern before the bearing replacement in October. Recommend inspection within 48 hours. Current job will finish fine." Predictive maintenance without a $50K platform.
A research prosthetic with KrillClaw embedded. It reads context from sensors and a wrist camera, then downloads fine-motor skill profiles from the cloud in real time. Picking up chopsticks for the first time? It loads a precision-grip pattern. Learning to play piano? Finger independence profile, downloaded in seconds. A prosthetics research lab testing new grasp patterns across dozens of patients? Same hardware, infinite configurations — each hand adapts to its wearer.
This is where edge AI is heading. KrillClaw is the runtime that makes it possible.
Layer adhesion drops on hour 6 of a 10-hour print. KrillClaw reads the temperature sensor, checks the filament spool weight (running low = thinner feed), and adjusts the flow rate mid-print. It texts you: "Print 80% done. Compensated for low spool — increased flow rate 5%. Should finish clean by 2am." You wake up to a perfect part instead of spaghetti.
Every example runs on a $3-5 chip over WiFi. The intelligence lives in the cloud. The agency lives on the device.
You already own incredible hardware. The electronics inside just never had a brain. A $3 KrillClaw chip and a WiFi connection brings the world's most capable AI models to any device — analyzing data, reacting in real time, taking initiative, running recurring actions. No replacement required.
Upgrade the brain, not the machine.
Your car's OBD-II port speaks data — but the car's own computer was frozen in time the day it shipped. Plug in a $3 ESP32 with KrillClaw. Now it reads engine codes, cross-references them with repair databases, monitors fuel efficiency trends, and texts you: "Your fuel economy dropped 8% this month — likely the air filter based on mileage. $12 part, 5-minute swap. Want me to order one?" A new car with these smarts costs $45K. This costs $3.
Your exercise bike has sensors for cadence, resistance, and heart rate — but its software is whatever the manufacturer shipped. Clip a KrillClaw node onto it. Now it reads your real-time performance, compares it to your training history, and adjusts coaching in context: "Your power output is down 15% from Tuesday — you might be under-recovered. Switching to an endurance zone ride instead of intervals today." A new smart bike with AI coaching: $3,000+. Retrofitting yours: $3.
Your CNC machine cost $50K but its controller is from 2015. Industrial predictive maintenance platforms cost $20K/year. Wire a KrillClaw chip to the vibration sensor and spindle temperature probe. It builds a baseline, detects anomalies, correlates with the maintenance log in flash memory, and messages the floor supervisor before things break. "Spindle bearing vibration trending up 12% since Monday — similar to the pattern before the October replacement. Recommend inspection within 48 hours." New CNC with smart diagnostics: $80K. Retrofitting: $3.
Your washer, dryer, pool pump, HVAC, water heater — all have sensors and control interfaces. None of them talk to each other or think for themselves. A KrillClaw chip on each one, connected to your home WiFi, and suddenly your house has a nervous system. The HVAC knows the dryer is about to dump heat and pre-adjusts. The pool pump shifts to off-peak rates. The water heater pre-heats before your usual shower time. A "smart home" retrofit from a contractor: $5K-15K. Doing it yourself with KrillClaw: $20 in chips.
Every device you own is an upgrade away. Not a replacement away.
Runs on a $3 chipWiFi, BLE, and serial built in. Connects to any LLM over the air.
← swipe to explore →
Sign up to get pre-flashed KrillClaw™ devices.
No spam. Just a ping when hardware ships.
git clone https://github.com/yoctoclaw/KrillClaw cd KrillClaw zig build -Doptimize=ReleaseSmall ./zig-out/bin/krillclaw "monitor sensors and alert on anomalies"