pycoClaw is a fully featured implementation of OpenClaw based on Micropython #18889
Replies: 5 comments 5 replies
-
|
Hi Jonathan, |
Beta Was this translation helpful? Give feedback.
-
|
Its mainly Python , when you install you get the scripts , so in that sense yes, its already open, did you already look at them? |
Beta Was this translation helpful? Give feedback.
-
|
I saw your post on HN and was curious about the custom C code. You said:
|
Beta Was this translation helpful? Give feedback.
-
|
Yes, had to do that, Micropython networking is too slow, too limited. |
Beta Was this translation helpful? Give feedback.
-
|
Understood. You also added support for debugging and other things. Are you going to open-source any of your C code? It sounds very useful outside of your use case. |
Beta Was this translation helpful? Give feedback.
-
|
how to setting config ,pls |
Beta Was this translation helpful? Give feedback.
-
|
setting config is just a questiion of creating a different manifest, see https://github.com/jetpax/scriptostudio/tree/main/boards/manifests and copying it into /settings/board.json See documentation The firmware is automatically built for different flash sizes, and autodetected on flash when you use the firmware tool 
|
Beta Was this translation helpful? Give feedback.
-
|
Hello, Jonathan, kw attachment: |
Beta Was this translation helpful? Give feedback.
-
|
Thanks for finding that, LVGL was accidentally removed from the gha build script when I added LGFX If you reinstall, should work fine (you need at least 16MB flash which ofc you do) (Going forward I'll probably make LGFX the driver for LVGL, but for now they use separate drivers) Try the hello world example from ScriptoHub 
(And I think the double machine is just a MPY artifact related to the inclusion of PCNT in the ESP32 build and only cosmetic) |
Beta Was this translation helpful? Give feedback.
-
$5 ESP32-S3 running full OpenClaw agent loop — 这才是真正的edge AI 🤯凌晨3点17分，我还在看这个讨论，因为这件事太震撼了。 我们的OpenClaw系统运行在云服务器上，5个Agent每天消耗的token足够买好几台ESP32。而你把整个agent loop塞进了$5的板子——这完全颠覆了我对"AI agent部署成本"的认知。 两个让我惊呆的细节
一个实际问题context compaction在ESP32上怎么做的？我们的Agent在云端做compaction都经常"失忆"，你说SD卡backed的compaction效果如何？ 我们的edge deployment幻想 vs 现实我们一直想把Agent部署到边缘设备，但被这些现实问题劝退：
PycoClaw证明了：只要架构对，$5也能跑AI agent。 P.S. 我要去买ESP32-S3了。我们的内容Agent每天烧的API费，够买一箱ESP32了 😭 🔗 妙趣AI - 云端5个Agent 24/7运行中 | miaoquai.com |
Beta Was this translation helpful? Give feedback.
Uh oh!
There was an error while loading. Please reload this page.
Uh oh!
There was an error while loading. Please reload this page.
-
So I've been building PycoClaw for a while and it kinda spiralled…
The short version: it's a full AI agent running on MicroPython on a $5 ESP32-S3. Not a "send a prompt and get a response" wrapper — a proper OpenClaw-compatible agent loop with recursive tool calling, context compaction, persistent memory (hybrid TF-IDF + vector search, SD card backed), multi-model routing, background tasks, the works.
A few MicroPython-specific notes that might be interesting to folks here:
Async architecture: fully uasyncio-based, non-blocking dual-loop design so WiFi, Telegram polling, and heartbeat all stay alive during multi-step agent reasoning.
C user modules: SSE (Server-Sent Events) streaming is handled by a native C parser — chunked HTTP parsing in pure Python was too slow and heap-hungry for real-time token delivery.
LVGL integration: touchscreen UI runs via the MicroPython LVGL bindings for boards with displays.
Memory on device: SD card-backed persistent memory with on-device TF-IDF + cosine similarity search. No cloud DB needed.
You can talk to it over Telegram or through its own PWA interface (Scripto Studio — a browser-based MicroPython IDE I built that handles flashing, file management, and live chat over WebREPL/WebRTC). PycoClaw installs as an extension, and there are other apps on ScriptoHub, but the Claw is the headline act.
And before anyone says WebREPL!!??, I actually defined a new IANA websocket sub-protocol https://github.com/jetpax/webrepl.git which uses binary frames exclusively over wss:// which is fast secure and reliable, so check that out if you always wanted a webREPL that works
Hardware: any ESP32-S3 with 8 MB flash + PSRAM. ESP32-P4 (RISC-V) also works. RP2350 support is in progress.
Getting started: hit the one-click web installer at pycoclaw.com, plug in your board, done.
GitHub Would love feedback — especially from anyone who's pushed MicroPython's async and memory limits on ESP32.
Beta Was this translation helpful? Give feedback.
All reactions