PicoClaw is a personal AI assistant written in Go that ships as a single binary for low-cost hardware. It was initiated by Sipeed, takes its design cues from NanoBot, and targets a core memory footprint under 10 MB and a boot time under one second on a 0.6 GHz single-core processor, so it can run on boards such as the LicheeRV-Nano as well as Raspberry Pi, servers, Docker, and Android phones (APK or Termux). Recent builds may use 10–20 MB while new features settle.
A Gateway process connects the agent to chat apps; a WebUI Launcher on port 18800 handles provider setup, channel setup, and chat from a browser, and the picoclaw CLI covers onboarding, interactive chat, model switching, cron jobs, skills, and MCP server entries. Roughly 95% of the core code was generated by an agent and refined through human review. The project is in early development and states that it should not be deployed to production before v1.0.
Builds target x86_64, ARM64, MIPS, RISC-V, and LoongArch. A model list in config.json maps names to provider-prefixed models, with rule-based routing that sends short queries to smaller models.
Features
- Chat channels: 19+ platforms including Telegram, Discord, WhatsApp, Weixin, QQ, Slack, Matrix, Delta Chat, DingTalk, Feishu/Lark, LINE, WeCom, VK, IRC, OneBot, MQTT, and MaixCam
- LLM providers: 30+ including OpenAI, Anthropic, DeepSeek, Doubao, Groq, Moonshot, MiniMax, Mistral, NVIDIA NIM, Cerebras, Xiaomi MiMo, Azure OpenAI, AWS Bedrock, GitHub Copilot, and local Ollama or vLLM endpoints, plus LiteLLM as a proxy
- MCP support: connect any Model Context Protocol server over stdio, SSE, or HTTP; picoclaw mcp add/list/test/edit/remove manages entries in the config
- Web search: DuckDuckGo built in, with Gemini Google Search, Baidu, Tavily, Brave, Kagi, Perplexity, SearXNG, GLM, and Sogou as configurable engines
- Skills: modular capabilities loaded from SKILL.md files, installable from ClawHub or GitHub registries
- Vision pipeline: images and files are base64-encoded and passed to multimodal models
- Scheduled tasks: one-time reminders, recurring tasks, and cron expressions through the cron tool
- Agent architecture: SubTurn sub-agent coordination, hooks (observers, interceptors, approval), steering messages into a running loop, and an event bus
- Security controls: a .security.yml policy, sensitive-data filtering, a sandbox, and cron gating for command jobs
- Deployment options: downloadable binaries, Docker Compose, a system tray UI on Windows and Linux, and an Android APK
