# Problem Addressed > PicoClaw is an ultra-lightweight personal AI assistant written in Go, designed to run on $10 hardware with less than 10MB RAM. Inspired by - Skill: `tools-only/problem-addressed-23` (Agent Skill, multi-file: 3 files) - Install (CLI): `npx skillmds@latest add tools-only/problem-addressed-23` - Raw SKILL.md: https://api.skillmd.com/api/skills/tools-only/problem-addressed-23/raw - Safety review: pending (external: skill-scanner PASS, skillspector PASS) - Works with: Claude Code, Claude.ai, OpenAI Codex - Category: Coding & Dev Tools - Author: tools-only (https://skillmd.com/u/tools-only) - Updated: 2026-09-29 - Page: https://skillmd.com/skills/tools-only/problem-addressed-23 --- # PicoClaw | Field | Value | | ------------- | ------------------------------------------------ | | Research Date | 2026-02-23 | | Primary URL | | | GitHub | | | Website | | | Version | v0.1.2 (latest at research date) | | License | MIT | | Company | Sipeed () | --- ## Overview PicoClaw is an ultra-lightweight personal AI assistant written in Go, designed to run on $10 hardware with less than 10MB RAM. Inspired by [nanobot](https://github.com/HKUDS/nanobot) and built via a self-bootstrapping process (95% agent-generated code), it provides a single-binary AI agent deployable on RISC-V, ARM, and x86 Linux devices — including decade-old Android phones via Termux — with 400× faster startup than comparable Python-based solutions. --- ## Problem Addressed | Problem | Solution | | -------------------------------------------------------- | --------------------------------------------------------------------------- | | AI assistants require expensive hardware (Mac Mini $599) | Runs on $9.9 LicheeRV-Nano RISC-V board; <10MB RAM in production | | Python/TypeScript agents have slow cold starts | Go binary boots in <1 second on a 0.6GHz single-core device | | AI infrastructure is hard to deploy at the edge | Single self-contained binary for RISC-V, ARM, x86; no external runtime | | Personal AI agents lack messaging channel integration | Telegram, Discord, QQ, DingTalk, LINE, WeCom built-in | | AI development lacks transparency about code provenance | 95% of core was agent-generated; self-bootstrapping methodology documented | | Setting up LLM providers is complex | Zero-code `model_list` config — add any OpenAI-compatible endpoint | --- ## Key Statistics | Metric | Value | Date Gathered | | ---------------- | ------------------- | ------------- | | GitHub Stars | 18,121 | 2026-02-23 | | GitHub Forks | 2,164 | 2026-02-23 | | Open Issues | 312 | 2026-02-23 | | Primary Language | Go | 2026-02-23 | | Repository Age | Since 2026-02-04 | 2026-02-23 | | Latest Release | v0.1.2 | 2026-02-23 | | Translations | 6 languages (README) | 2026-02-23 | --- ## Key Features ### Runtime and Performance - **<10MB RAM**: Core functionality fits in under 10MB; recent PR merges may push to 10–20MB until planned optimization pass - **<1 second startup**: Cold-start on 0.6GHz single-core hardware - **Single binary**: Cross-compiled for RISC-V, ARM64, x86\_64; no Node.js, Python, or JVM - **Docker Compose support**: Optional containerized deployment with gateway and agent profiles - **Android/Termux support**: Runs on old Android phones via Termux with `proot` ### Multi-Channel Messaging Six built-in channel integrations with unified `allow_from` allowlist security: - **Telegram** — token-based bot; recommended for personal use - **Discord** — bot token + MESSAGE CONTENT INTENT; optional mention-only mode - **QQ** — AppID + AppSecret via QQ Open Platform - **DingTalk** — Client ID + Client Secret - **LINE** — Channel Secret + Channel Access Token + HTTPS webhook - **WeCom** — both Bot (webhook URL) and App (CorpID + AgentID) modes ### AI Provider Configuration Zero-code multi-provider setup via `model_list` array in `config.json`: - OpenRouter, Anthropic, OpenAI, Google Gemini, Zhipu — any OpenAI-compatible endpoint - Per-model `api_key`, `max_tokens`, and `temperature` settings - Named model aliases (e.g., `"gpt4"`, `"claude-sonnet-4.6"`) referenced by agent defaults ### Agent Workflows - **Full-Stack Engineer**: Code generation and file operations - **Logging & Planning Management**: Scheduling and memory-backed planning - **Web Search & Learning**: Tavily, Brave Search, DuckDuckGo with auto-fallback ### Edge Hardware Targets Purpose-built for Sipeed's own hardware product line: - `$9.9` LicheeRV-Nano (RISC-V, Ethernet or WiFi6) — minimal home assistant - `$30–$50` NanoKVM / `$100` NanoKVM-Pro — automated server maintenance - `$50` MaixCAM / `$100` MaixCAM2 — smart camera monitoring --- ## Technical Architecture ### Stack Components | Component | Technology | | ---------------- | ------------------------------------------------------ | | Core Language | Go (single binary, `make build` / `make build-all`) | | Config | JSON (`~/.picoclaw/config.json`) | | Web Search | Tavily, Brave Search, DuckDuckGo (auto-fallback) | | Gateway | HTTP webhook server (channels: Telegram, Discord, etc) | | Agent Mode | One-shot (`-m "..."`) or interactive | | Build | Makefile (`make deps`, `make build`, `make install`) | | Container | Docker Compose with `gateway` and `agent` profiles | ### Data Flow (Agent Message Handling) 1. Inbound message via channel (Telegram, Discord, CLI, etc.) 2. `allow_from` allowlist check — deny if sender not in list 3. Agent loop: LLM provider call with context + tools 4. Tool execution (web search, file ops, code run) 5. Response dispatched back to originating channel ### AI-Bootstrapped Development PicoClaw was built using its own predecessor (nanobot/OpenClaw) to drive the Go migration: - 95% of core code is agent-generated - Human-in-the-loop refinement for architecture decisions - Self-bootstrapping methodology serves as a living demo of agentic software development --- ## Installation & Usage ### Precompiled Binary (Android/Termux example) ```bash wget https://github.com/sipeed/picoclaw/releases/download/v0.1.2/picoclaw-linux-arm64 chmod +x picoclaw-linux-arm64 pkg install proot termux-chroot ./picoclaw-linux-arm64 onboard ``` ### Build from Source ```bash git clone https://github.com/sipeed/picoclaw.git cd picoclaw make deps make build # single platform make build-all # cross-compile all platforms make install ``` ### Docker Compose ```bash cp config/config.example.json config/config.json # Edit config.json: set API keys, model_list, channels docker compose --profile gateway up -d docker compose run --rm picoclaw-agent -m "What is 2+2?" ``` ### Quick Start CLI ```bash # 1. Initialize picoclaw onboard # 2. Configure ~/.picoclaw/config.json with model_list and API keys # 3. Chat picoclaw agent -m "What is 2+2?" # 4. Start gateway (messaging channels) picoclaw gateway ``` ### Minimal `config.json` ```json { "agents": { "defaults": { "workspace": "~/.picoclaw/workspace", "model": "claude-sonnet", "max_tokens": 8192, "temperature": 0.7, "max_tool_iterations": 20 } }, "model_list": [ { "model_name": "claude-sonnet", "model": "anthropic/claude-sonnet-4-6", "api_key": "your-anthropic-key" } ], "tools": { "web": { "duckduckgo": { "enabled": true, "max_results": 5 } } } } ``` --- ## Relevance to Claude Code Development ### Applications 1. **Edge Deployment Reference**: PicoClaw proves a Go-native AI agent can run Claude-class LLMs on $10 RISC-V hardware. Relevant when designing Claude Code plugins that must run in resource-constrained CI environments or embedded contexts. 2. **AI-Bootstrapped Methodology**: The self-bootstrapping Go migration (95% agent-generated core) is a concrete example of using Claude Code agents to autonomously rewrite an existing codebase in a new language — a pattern directly applicable to large-scale refactoring tasks. 3. **Zero-Code Provider Config**: The `model_list` JSON array pattern for adding AI providers without code changes is a clean configuration pattern for skills that need to support multiple LLM backends. 4. **Multi-Channel Gateway Architecture**: The unified `allow_from` + channel config pattern is worth studying for any Claude Code plugin that needs to dispatch AI responses to multiple messaging platforms. ### Patterns Worth Adopting 1. **Single-binary cross-compilation**: PicoClaw's `make build-all` with goreleaser demonstrates how to ship a Go agent binary for ARM/RISC-V/x86 from a single CI job — applicable if Claude Code plugins ever ship native helper binaries. 2. **Named model aliases in config**: Referencing models by alias (`"gpt4"`, `"claude-sonnet"`) rather than full provider strings keeps agent config portable across provider changes. 3. **Auto-fallback web search**: The DuckDuckGo → Tavily → Brave Search fallback chain is a practical pattern for tool reliability without hard provider dependencies. 4. **Termux deployment pattern**: `pkg install proot && termux-chroot ./binary onboard` shows how to run a Linux binary on Android without root — useful for documenting install paths in skills targeting mobile edge devices. ### Competitive Context PicoClaw sits between nanobot (Python, >100MB RAM) and ZeroClaw (Rust, <5MB RAM): | Dimension | NanoBot (Python) | PicoClaw (Go) | ZeroClaw (Rust) | | ---------------- | ---------------- | ------------- | --------------- | | RAM | >100MB | <10–20MB | <5MB | | Startup (0.8GHz) | >30s | <1s | <10ms | | Language | Python | Go | Rust | | Cost hardware | ~$50 SBC | $10 SBC | $10 SBC | | Stars (Feb 2026) | — | 18,121 | 14,966 | | Channels | Limited | 6+ | 15+ | | License | MIT | MIT | Other | ZeroClaw has more channels, more providers, and lower memory usage. PicoClaw has more community traction (18K vs. 15K stars) and is backed by Sipeed's hardware ecosystem. See [zeroclaw.md](./zeroclaw.md) for the Rust counterpart. ### Caution: Very Early Stage PicoClaw was created on 2026-02-04 and hit 18K stars within two weeks. The project is in early development with known network security issues — the README explicitly warns against production deployment before v1.0. Recent PR merges have increased RAM usage to 10–20MB; resource optimization is scheduled post feature-stabilization. Treat as reference architecture, not production infrastructure. --- ## References | Source | URL | Accessed | | ---------------------- | ---------------------------------------------------------------- | ---------- | | GitHub Repository | | 2026-02-23 | | GitHub README | | 2026-02-23 | | Official Website | | 2026-02-23 | | Sipeed Company | | 2026-02-23 | | GitHub API (repo meta) | `gh api repos/sipeed/picoclaw` | 2026-02-23 | | GitHub API (release) | `gh api repos/sipeed/picoclaw/releases/latest` | 2026-02-23 | | ROADMAP.md | | 2026-02-23 | **Research Method**: Information gathered from GitHub web page, GitHub API (stars, forks, issues, releases), and project README. Repository created 2026-02-04; rapidly growing community. --- ## Freshness Tracking | Field | Value | | ------------------ | ------------------------- | | Version Documented | v0.1.2 | | Release Date | 2026-02-23 (approximate) | | GitHub Stars | 18,121 (as of 2026-02-23) | | GitHub Forks | 2,164 (as of 2026-02-23) | | Next Review Date | 2026-05-23 | **Review Triggers**: - v1.0 release (README flags this as the production-readiness milestone) - Stars exceed 30K (sustained community traction indicator) - RAM footprint optimization milestone completed - Security advisories (project flagged active network security issues pre-v1.0) - New channels or provider integrations added - Changes to Sipeed hardware lineup that affect deployment targets