PicoClaw
Quick Facts
- Category
- Agent runtime (not a model)
- Language
- Go
- Memory footprint
- Under 10MB RAM at runtime
- Partner
- Sipeed (RISC-V hardware maker)
- Target hardware
- $10 RISC-V SBCs · IP cameras · routers · microcontrollers · embedded Linux
- Architecture
- Agent-on-device; inference off-device via MQTT / gRPC
- Inference
- Offloaded to edge server (OpenClaw host) or cloud endpoint
- License
- Open source (see Claw ecosystem on GitHub)
Summary
PicoClaw is the IoT and embedded member of the Claw family. It was developed in partnership with Sipeed, the Shenzhen-based company behind the sub-$10 RISC-V boards (LicheePi, MaixCube) that have made real embedded Linux cheap enough for one-per-sensor deployments. The technical constraint — under 10MB RAM at runtime — rules out running an LLM locally. PicoClaw's architectural premise is that the agent protocol can live on the device while the inference lives elsewhere.
The on-device runtime handles: receiving events (MQTT, serial, GPIO), dispatching tool calls (local GPIO writes, sensor reads, actuator control), maintaining a minimal state machine, and forwarding LLM calls via MQTT or gRPC to an OpenClaw or ExoClaw host. The agent looks the same from the control plane's perspective; the inference just happens a network hop away.
Architecture
- Device runtime — Go binary. Single static executable; deploys via scp or an OTA channel.
- Local tools — GPIO, I²C, SPI, UART, camera / microphone capture. Tool definitions exposed in the common Claw tool schema.
- Transport — MQTT for fleet deployments, gRPC for point-to-point. TLS for both.
- Upstream — OpenClaw host or ExoClaw cloud endpoint. The upstream runs the actual agent loop, sees PicoClaw's tools as remote tools, and dispatches back through the same transport.
- State — minimal. PicoClaw keeps the last N events in memory and optionally flushes to local flash. Canonical state lives upstream.
- Security — device attestation, per-device credentials, FrawdBot analysis of all tool calls at the upstream boundary.
When to Choose PicoClaw
PicoClaw is the right runtime when the agent's environment is the physical world and the hardware envelope is small: smart cameras with agentic triggers, industrial sensor fleets, retail IoT, home automation, agricultural deployments. Any scenario where one agent per device times tens-to-thousands of devices makes more sense than a few centralized agents polling.
Tradeoffs
- Network dependency. PicoClaw needs a path to an inference upstream. Fully air-gapped devices need a different architecture (or a much smaller on-device model via llama.cpp — but that exceeds PicoClaw's footprint).
- Latency floor. Round trips to the inference upstream add latency. Real-time control loops (sub-100ms) should do the control logic on-device and use the LLM asynchronously.
- Fleet management overhead. A thousand devices is a thousand identities to manage, update, and monitor. PicoClaw provides OTA hooks; ops cost is real.
- Limited local intelligence. Without the LLM on-device, PicoClaw can't make free-form decisions offline. Design for graceful degradation.
Related
- ZeroClaw — similar-footprint runtime with on-device inference support.
- OpenClaw — the typical inference upstream.
- ExoClaw — the typical cloud upstream.