PycoClaw
by PycoClaw (independent open-source project by Jonathan Peace)
A full OpenClaw agent running on a $5 microcontroller
PycoClaw is an open-source MicroPython build of the OpenClaw agent framework that runs a complete tool-calling AI agent on an ESP32 microcontroller costing roughly five dollars. It gives embedded and robotics developers on-device reasoning, persistent memory and direct GPIO, CAN and I2C hardware control, with the agent able to write and hot-load its own scripts at runtime.
PycoClaw is a MicroPython implementation of the OpenClaw agent framework that runs entirely on a microcontroller, putting a full recursive tool-calling agent loop on hardware costing about five dollars rather than on a server. Created by Jonathan Peace and released under the MIT licence in early 2026, it targets ESP32-S3, ESP32-P4 and ESP32-C6 in production with Raspberry Pi RP2350 support still in progress, and requires at least 8MB of flash and 4MB of PSRAM; the firmware image itself is around 2MB. The project is substantially larger than that footprint suggests: an agent core of roughly 26,000 lines, about 18,000 lines of custom C extensions written to make MicroPython fast enough, a browser-based progressive web app IDE called Scripto Studio that handles chat, file editing and debugging, and ScriptoHub, a curated repository of hardware Skills with malware checking. Architecturally it runs a fully uasyncio dual-loop design so Wi-Fi, Telegram polling and heartbeats stay alive during multi-step reasoning, keeps persistent hybrid memory that combines TF-IDF keyword search with vector embeddings backed by an SD card, and routes across LLM providers using an OpenAI-compatible specification with tolerant tool-argument coercion so GLM, Qwen, Moonshot, Gemini and Ollama endpoints all work behind failover. Its distinguishing capability is self-programming: the agent writes and hot-loads its own MicroPython scripts and performance-critical C extensions at runtime, so a device can adapt its own firmware in the field — reverse-engineering an unknown motor protocol, for example — without a rebuild. It also supports cron scheduling, subagent spawning, over-the-air updates, one-click browser flashing, and direct GPIO, CAN, I2C, SPI and LVGL touchscreen control that server-hosted OpenClaw has no path to.
An embedded or robotics engineer who wants an LLM agent making decisions on the device itself — with GPIO, CAN and I2C in reach — rather than round-tripping every decision through a server.
A complete tool-calling agent with persistent memory and the ability to rewrite its own firmware scripts, running on a roughly $5 ESP32-S3.
At a Glance
- Category
- Agent Development
- Pricing
- Free
- Target Market
- Embedded Systems Engineers, Robotics Engineers, IoT Developers, Hardware Prototypers
- Deployment
- Self-hosted, Open-source, Edge-first
- Founded
- 2026
Key Features
- ✓Full agent loop on-device
Recursive tool calling, context compaction and subagent spawning inside roughly 2MB of microcontroller firmware, with no server in the path.
- ✓Self-programming runtime
The model writes and hot-loads its own MicroPython scripts and C extensions, so device behaviour changes without a firmware rebuild.
- ✓Hybrid persistent memory
TF-IDF keyword search combined with vector embeddings, backed by SD card, so conversation and context survive reboots.
- ✓Multi-provider LLM routing
An OpenAI-compatible layer with tolerant tool-argument coercion lets GLM, Qwen, Moonshot, Gemini and Ollama endpoints be swapped with failover.
- ✓Direct hardware control
GPIO, CAN, I2C, SPI and LVGL touchscreen drivers are first-class tools the agent can call, which server-hosted OpenClaw cannot do.
- ✓Scripto Studio browser IDE
A progressive web app that flashes firmware in one click then serves as chat client, file editor and debugger, with no local toolchain.
- ✓Dual-loop uasyncio architecture
Wi-Fi, Telegram polling and heartbeat stay responsive throughout multi-step agent reasoning instead of blocking on the agent loop.
Capabilities
Use Cases
- •Self-adapting robotics
A robot rewrites its own control scripts, for example reverse-engineering an unknown motor protocol by reasoning about its responses.
- •Battery-powered wearable assistants
A pocket device holds personal context in persistent memory and answers over Telegram without needing a phone app.
- •Field IoT diagnostics
A sensor node reasons about anomalous readings on site and reports conclusions rather than shipping raw telemetry to a cloud pipeline.
- •Hardware bring-up and prototyping
Engineers drive GPIO, CAN and I2C from natural language during bring-up, then keep the generated script as production code.
- •Embedded agent research
A cheap MIT-licensed platform for studying how agents behave under hard memory, power and latency constraints.
Ideal For
Best For
- ✓Putting a conversational agent on a battery-powered ESP32 device without a companion server
- ✓Robotics and hardware bring-up where the agent needs direct GPIO, CAN and I2C access
- ✓Self-adapting IoT firmware that rewrites its own scripts in the field
- ✓Research and hobbyist work on agent behaviour under hard memory and latency constraints
- ✓Prototyping chat-driven hardware over Telegram or WebRTC with no local toolchain to install
Not Ideal For
- ✗Enterprises that need a vendor, an SLA or any compliance posture — this is a solo-maintained MIT project with none of them
- ✗Air-gapped or offline deployments: nothing is inferred on-device, so no network means no agent
- ✗Teams standardised on RP2040 or low-memory ESP32 boards, which the 8MB flash / 4MB PSRAM floor excludes, with RP2350 support still unfinished
- ✗Safety-critical or hard-real-time control, where a non-deterministic model in the loop and interpreted execution are each disqualifying
Integrations
Deployment
Market Analysis
Pros
- ✓Genuinely delivers a full agent loop on about $5 of silicon — MIT licensed, one-click browser flashing, no toolchain to install
- ✓Self-programming is a real capability rather than a demo: the agent writes MicroPython and C extensions at runtime, so a device can adapt its own firmware in the field
- ✓Broader than its embedded peers on channels and providers — Telegram, Scripto Studio and WebRTC, with OpenAI, Gemini, Ollama, GLM, Qwen and Moonshot behind a tolerant OpenAI-compatible layer
- ✓Picked up by the embedded press within weeks (CNX Software, Adafruit) and surfaced on Hacker News, which is unusual reach for a solo project
- ✓Non-blocking dual-loop design means Wi-Fi and chat stay alive during multi-step reasoning, which is the failure mode most naive embedded agent ports hit
Cons
- ✗The open-source claim is only partly delivered — CNX Software reported the firmware source is not actually published on GitHub despite the MIT label, only the website source, so you cannot currently audit or rebuild what you flash
- ✗Nothing is inferred on-device: every agent turn needs network access and a paid LLM API key, so it is neither an offline nor an air-gapped solution and running cost scales with usage
- ✗The hardware floor of 8MB flash plus 4MB PSRAM rules out most commodity ESP32 boards and all RP2040 devices, and RP2350 support is still listed as in progress
- ✗Interpreted execution is slower than purpose-written C, raised directly in the Hacker News thread; the maintainer's answer is that the model generates C extensions for hot paths, which is a mitigation rather than parity
- ✗A single-maintainer project at roughly 153 GitHub stars with no company behind it: no support contract, no SLA, no security review and no compliance posture
- ✗Supports far fewer chat channels than upstream OpenClaw — two against fourteen or more — so it is not a drop-in replacement for a server deployment
Pricing
Open source (MIT)
$0
- ✓Full agent framework under the MIT licence
- ✓Scripto Studio browser IDE
- ✓ScriptoHub skills repository
- ✓One-click web flashing
- ✓Over-the-air updates
PycoClaw itself is free under the MIT licence, with no vendor, subscription, seat count or support contract, and nothing gated behind a paid tier. The real cost sits elsewhere: a compatible board, roughly $5 for an ESP32-S3 with at least 8MB flash and 4MB PSRAM, plus per-token spend with whichever LLM provider you point it at, because no inference happens on-device. The maintainer's own example configuration exposes a monthly spend throttle with a $5 budget in it, which is a fair signal that token cost rather than hardware is the recurring line item. Because provider routing is OpenAI-spec compatible, a self-hosted Ollama endpoint can replace the paid API and drop marginal cost to electricity, at the price of running that server yourself.
Security & Compliance
Connect
Sources
This page was written from 4 sources, 3 on domains other than pycoclaw.com.
Stay Ahead of the Curve
Weekly enterprise AI insights for technology leaders. No spam, no vendor pitches—unsubscribe anytime.
SubscribeRelated Products
Sim
Open-source, Apache-2.0 AI workspace for building, deploying and governing AI agent workflows
Zep
Context layer for AI agents: temporal knowledge graphs that give agents governed memory of users, data and documents
Restate
Durable execution runtime that keeps AI agents and backend workflows running through crashes and failures
Mastra
Open-source TypeScript framework and platform for building, deploying and observing production AI agents and workflows