picoclaw

PicoClaw is an independent open-source personal AI assistant written entirely in Go, initiated by Sipeed. It is designed for extremely constrained hardware: the project claims a core memory footprint under 10MB and boot times under one second on low-power single-core processors, with a single binary targeting x86_64, ARM64, MIPS, RISC-V, and LoongArch. The README notes that recent builds may use 10-20MB RAM due to rapid merges, with optimization planned. The assistant supports a broad range of LLM providers through a model_list configuration using a protocol/model format, including OpenAI, Anthropic, Google Gemini, OpenRouter, Zhipu, DeepSeek, Volcengine, Qwen, Groq, Moonshot, Minimax, Mistral, NVIDIA NIM, Cerebras, NEAR AI Cloud, Novita, Xiaomi MiMo, Ollama, vLLM, LiteLLM, Azure OpenAI, GitHub Copilot, Antigravity, and AWS Bedrock. Local deployment via Ollama and vLLM is documented. Chat integration covers 19+ messaging platforms such as Telegram, Discord, WhatsApp, Weixin, QQ, Slack, Matrix, Delta Chat, DingTalk, Feishu/Lark, LINE, WeCom, VK, IRC, OneBot, MQTT, MaixCam, and Pico clients. Webhook-based channels share a Gateway HTTP server. Other documented capabilities include native Model Context Protocol (MCP) server integration, a vision pipeline that base64-encodes images and files for multimodal models, rule-based model routing to send simple queries to lightweight models, web search via DuckDuckGo, Gemini Google Search, Baidu, Tavily, and Brave, and a WebUI launcher for browser-based configuration and chat. Installation options include precompiled binaries, building from source with Go 1.25+ and Node.js/pnpm for the frontend, Docker Compose, Android APK, and Termux. The project is MIT licensed and explicitly warns it is in early development and not ready for production before v1.0.