PicoClaw: The Go AI Assistant That Boots in Under a Second on $10 Hardware
Tiny, Fast, and Deployable anywhere, automate the mundane, unleash your creativity.
At a glance
- What is it?
- PicoClaw is an open-source Go AI assistant that runs on RISC-V, ARM, MIPS, and x86 hardware with under 10 MB of RAM, delivering MCP tool support, multi-channel deployment, and a vision pipeline in a single static binary. It was built by Sipeed and targets edge devices, embedded boards, and developers who cannot afford the memory overhead of Node.js or Python runtimes.
- Who is it for?
- PicoClaw is the right choice for developers who need an AI assistant on constrained hardware such as RISC-V boards, ARM single-board computers, or Android phones, and who need MCP tool integration and multi-channel deployment without a Node.js or Python runtime. The README explicitly warns against production deployment before v1.0 because recent PR merges have pushed RAM usage to 10-20 MB in some builds and security issues may remain unresolved.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 5 days ago.
- What is it written in?
- Mainly Go, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 29, 2026, and from our analysis. They are not legal advice.
Editorial analysis
What PicoClaw is and the hardware it targets
PicoClaw is an AI assistant written entirely in Go, described in the README as "ultra-lightweight" and targeting hardware as inexpensive as $10. The project originated at Sipeed, a company known for RISC-V hardware products, and the README notes that a LicheeRV-Claw board is available for purchase as a reference deployment target. The core claim is that PicoClaw boots in under one second on a 0.6 GHz single-core processor and runs with a memory footprint under 10 MB. The README compares it directly to NanoBot (Python-based, over 100 MB RAM) and OpenClaw (TypeScript-based, over 1 GB RAM, typically requiring a Mac Mini class machine). PicoClaw is not a fork of either; the README explicitly states it was "written entirely in Go from scratch".
The self-bootstrapping origin and Go architecture
The README describes the build process as "self-bootstrapping": the AI Agent itself drove the architecture migration and code optimization from an initial concept to the Go implementation. The README claims 95% of core code was generated by an agent through human-in-the-loop review. The go.mod file sets the module path as `github.com/sipeed/picoclaw` and requires Go 1.25.13. The project builds a single static binary with `CGO_ENABLED=0` and build tags `goolm,stdjson`. The static binary design is key to the portability claim: one binary runs on RISC-V, ARM, MIPS, and x86 without runtime dependencies. The `cmd/picoclaw/` directory contains the main entry point.
Configuring PicoClaw with environment variables
Configuration follows a twelve-factor app model using a `.env` file. The repository includes a `.env.example` with the available settings grouped by function:
# ANTHROPIC_API_KEY=sk-ant-xxx
# OPENAI_API_KEY=sk-xxx
# GEMINI_API_KEY=xxxChannel configuration for Telegram:
# TELEGRAM_BOT_TOKEN=123456:ABC...For timezone:
TZ=Asia/ShanghaiThe `.env.example` lists providers including OpenRouter, Zhipu, Anthropic, OpenAI, Gemini, and ModelScope. Feishu (Lark), Discord, WeChat, and WeCom channels are configured through similar token-based environment variables. Optional Brave Search integration uses `BRAVE_SEARCH_API_KEY`. The v0.2.5 release added reading timezone from the TZ or ZONEINFO environment variable, so the `TZ` entry in the example is used by the runtime.
Supported channels and the LLM provider list
The go.mod dependencies reveal the full list of supported messaging channels: Telegram (via the mymmrac/telego library), Discord (bwmarrin/discordgo), Slack (slack-go/slack), LINE (line-bot-sdk-go), VK (SevereCloud/vksdk), Lark/Feishu (larksuite/oapi-sdk-go), WhatsApp (go.mau.fi/whatsmeow), Matrix (ergochat/irc-go), and WeChat. An MQTT channel is also present via the eclipse/paho.mqtt.golang dependency, which is not commonly found in AI assistant frameworks and enables IoT integration scenarios. The DingTalk channel uses the open-dingtalk/dingtalk-stream-sdk-go library. The LLM provider list includes Anthropic (anthropics/anthropic-sdk-go), OpenAI-compatible endpoints (openai/openai-go), AWS Bedrock (aws-sdk-go-v2/service/bedrockruntime), Azure (azure-sdk-for-go), Kimi, Minimax, Xiaomi MiMo, and ModelScope. Kagi search is available via kagisearch/kagi-openapi-golang. v0.2.1 added model routing: simple queries route to lightweight models automatically to save API costs. v0.2.9 added configurable Sogou-backed web search and MCP server management in the Web UI.
MCP integration and the agent architecture
PicoClaw implements the Model Context Protocol (MCP) using the official Go SDK (`modelcontextprotocol/go-sdk v1.6.1`). MCP allows connecting any MCP server to extend agent capabilities: tools, resources, and prompts defined by external MCP servers become available to the agent without modifying PicoClaw's core. v0.2.8 added MCP CLI commands: `show`, `add`, `list`, `remove`, `test`, and `edit`. v0.2.9 extended MCP management to the Web UI. The agent architecture as of v0.2.4 introduced SubTurn, Hooks, Steering, and EventBus components. Hooks allow custom logic at specific points in the agent turn cycle, including a respond action added in v0.2.6. The `isolation` flag, also added in v0.2.6, controls agent process separation. The vision pipeline, added in v0.2.1, accepts images and files sent directly to the agent and performs automatic base64 encoding for multimodal LLM calls. The go.mod also includes `github.com/pion/webrtc/v3` and `github.com/pion/rtp`, indicating real-time media handling capability alongside the standard text and image paths. The JSONL memory store, introduced in v0.2.1, gives the agent persistent memory across sessions.
Security caveats and the production warning
The README contains a prominent caution block with several warnings. PicoClaw has not issued any official tokens or cryptocurrency; any claims on pump.fun or similar platforms are described as scams. The README states the only official website is picoclaw.io and the company website is sipeed.com, noting that many `.ai`, `.org`, `.com`, and `.net` domains have been registered by third parties. More directly relevant to deployment: "PicoClaw is in early rapid development. There may be unresolved security issues. Do not deploy to production before v1.0." The v0.2.4 release added a `.security.yml` configuration file and sensitive data filtering as part of security hardening, but the project itself acknowledges that rapid development makes comprehensive security review difficult.
Limitations: RAM creep and Android availability
The sub-10MB RAM claim now carries a footnote in the README: "Recent builds may use 10-20MB due to rapid PR merges. Resource optimization is planned after feature stabilization." This is an honest acknowledgment that the headline number no longer holds for the latest builds. Teams deploying on the smallest RISC-V boards (0.5-1 GB total RAM) are not significantly affected by the 10-20 MB range, but the trend shows that memory usage grows with features. The README notes Android support was added on 2026-03-31, with APK downloads at picoclaw.io/download. Docker Compose and Web UI Launcher support arrived with v0.2.0. The system tray UI for Windows and Linux was added in v0.2.3. For teams wanting a mature, production-ready lightweight agent framework, there is no fully air-gapped alternative at this weight class; the Python-based NanoBot, which the README cites at github.com/HKUDS/nanobot, uses over 100 MB RAM and requires a more capable host.
Maintenance, community, and the MIT licence
PicoClaw is released under the MIT licence, allowing commercial use and redistribution. The last push was on 2026-09-24. The most recent versioned release was v0.3.1 on 2026-07-03, with v0.2.9 released on 2026-05-29. The nightly build was last updated on 2026-07-02. The project tracks community involvement: the README notes 26,000 stars reached by March 2026 and active community maintainer roles launched with v0.2.3. Coordination happens on Discord at discord.gg/V4sAZ9XWpN and via WeChat (contact details in the repository assets). The hardware reference platform is the LicheeRV-Claw from Sipeed, available on AliExpress. The ROADMAP.md file in the repository documents future directions. The Makefile includes an `integration-test` target and a `lint-docs` target, indicating the project maintains CI coverage beyond unit tests. The `.golangci.yaml` file in the root configures Go static analysis linting. The `.goreleaser.yaml` file manages cross-platform binary release builds. The `workspace/` directory in the root holds the default agent workspace configuration that ships with the binary, and `onboard_workspace_embed.go` embeds it at compile time.
Editorial conclusion
PicoClaw is the right choice for developers who need an AI assistant on constrained hardware such as RISC-V boards, ARM single-board computers, or Android phones, and who need MCP tool integration and multi-channel deployment without a Node.js or Python runtime. The README explicitly warns against production deployment before v1.0 because recent PR merges have pushed RAM usage to 10-20 MB in some builds and security issues may remain unresolved. Verify the actual binary size and RAM footprint for your target board against the current release before deploying. The last push was on 2026-09-24 and the most recent versioned release was v0.3.1 on 2026-07-03.
Frequently asked questions
What is PicoClaw?
PicoClaw is an open-source Go AI assistant designed to run on minimal hardware, including $10 RISC-V and ARM boards. It supports MCP tool integration, multi-channel deployment (Telegram, Discord, WhatsApp, and others), a vision pipeline for image inputs, and model routing to select lightweight models for simple queries.
Is PicoClaw better than OpenClaw?
The README's comparison table claims PicoClaw uses under 10 MB RAM (versus over 1 GB for OpenClaw) and boots in under 1 second on a 0.6 GHz core (versus over 500 seconds). It runs on $10 hardware rather than requiring a Mac Mini class machine. The README notes that recent builds may use 10-20 MB due to rapid PR merges, and resource optimization is planned after feature stabilization.
Is PicoClaw open source?
Yes. The repository is public at github.com/sipeed/picoclaw under the MIT licence. The README states it is an independent open-source project initiated by Sipeed, written entirely in Go from scratch.
Official sources
Add this badge to your README
If you maintain this project, the badge below links readers to this analysis and shows its maintenance status from the daily GitHub snapshot. Paste the markdown into your README; add ?metric=license or ?metric=stars to the image URL for a different field.
[](https://hysenlabs.com/projects/sipeed-picoclaw)