Key takeaways
- World's first AI assistant on a $5 microcontroller — pure C, no OS, 0.5W power
- Local-first memory survives reboots, stored on flash — no cloud dependency
- Grew to ~5.5K stars by June 2026, but commits stalled after April — velocity has cooled since the February hype wave
FAQ
What is MimiClaw?
An AI assistant that runs on a $5 ESP32-S3 chip. No Linux, no Node.js — just pure C, USB power, and Telegram control.
How much does MimiClaw cost?
~$10 for hardware (ESP32-S3 board). Claude or OpenAI API costs for LLM.
Who competes with MimiClaw?
PicoClaw (Go, RISC-V), NullClaw (Zig, minimal), BabyClaw (VPS + Telegram).
Is MimiClaw still maintained?
The repo is not archived and reached ~5.5K stars, but the last release (v0.1.1) shipped March 17, 2026 and the last push was April 21, 2026 — development has slowed noticeably as of June 2026.
Executive Summary
MimiClaw is the most extreme OpenClaw alternative: it runs on a $5 ESP32-S3 microcontroller. No Linux. No Node.js. No VPS. Just pure C firmware, 0.5W power consumption, and Telegram control. Plug it into USB power, connect to WiFi, and talk to it — it handles tasks and remembers everything on local flash storage. The project is run by MemoV, Inc., which has a hardware-kit waitlist at mimiclaw.io; funding is not publicly disclosed.
Status (as of June 2026): Active but slowing. The repo is not archived and sits at ~5.5K stars (up from 2.8K in February), but only two releases have shipped — v0.1.0 (Feb 9, 2026) and v0.1.1 (Mar 17, 2026) — and the last push to the repo was April 21, 2026.
| Attribute | Value |
|---|---|
| Language | C |
| License | MIT |
| GitHub Stars | 5.5K ★ (Jun 2026) |
| Latest Release | v0.1.1 (Mar 17, 2026) |
| Hardware | ESP32-S3 (~$5-10) |
| Power | 0.5W USB |
Key Capabilities
| Capability | Description |
|---|---|
| $5 hardware | ESP32-S3 with 16MB flash, 8MB PSRAM |
| No OS | Pure C firmware, no Linux/RTOS |
| USB powered | 0.5W, runs 24/7 on any USB port |
| Telegram control | Message it from anywhere |
| WebSocket gateway | LAN clients via port 18789 |
| Local memory | SOUL.md, USER.md, MEMORY.md on flash |
| Multi-provider | Anthropic Claude or OpenAI GPT, switchable at runtime |
| OTA updates | Firmware updates over the air (v0.1.1) |
| Tool calling | ReAct agent loop with web search, cron |
| Cron scheduler | AI can create its own scheduled tasks |
| Heartbeat | Autonomous task checking from HEARTBEAT.md |
Memory Files
| File | Purpose |
|---|---|
| SOUL.md | Bot personality |
| USER.md | Your info and preferences |
| MEMORY.md | Long-term memory |
| HEARTBEAT.md | Autonomous task list |
| cron.json | Scheduled jobs (AI-created) |
| 2026-02-21.md | Daily notes |
Tools
- web_search — Tavily or Brave Search API
- get_current_time — Sync system clock
- cron_add/list/remove — AI manages its own schedule
Strengths
- Cheapest hardware — $5-10 for the entire device
- Lowest power — 0.5W, runs continuously on USB
- True portability — Thumb-sized, pocket-friendly
- Local-first — All data on device flash
- Self-scheduling — AI creates its own cron jobs
- Heartbeat autonomy — Proactively acts on task list
Cautions
- Slowing development — Last release v0.1.1 (Mar 17, 2026); last push Apr 21, 2026; 106 open issues as of June 2026 (GitHub)
- API costs add up fast — XDA's reviewer hit 50% of token budget in an hour with the hardcoded Opus model; switching to Haiku is recommended (XDA Developers)
- Limited compute — ESP32-S3 is a microcontroller, not a computer
- C expertise needed — Flashing firmware requires ESP-IDF v5.5 knowledge
- Telegram-first — Only Telegram plus a LAN WebSocket gateway; no Discord, Slack, or WhatsApp
- WiFi required — No offline LLM capability
- No sandboxing — Limited security isolation
What Developers Say
From XDA Developers' hands-on review (Samir Makwana, April 2026):
"The response times of about 3 to 8 seconds for simple questions are quite acceptable."
"After running MimiClaw on ESP32-S3 for an hour, I wanted to check the token usage. To my shock, it was already at 50 percent."
"This isn't a plug-and-play style product. I had to build and flash the firmware."
On Hacker News, MimiClaw threads drew little traction (single-digit points); the most-voted reply was skepticism of the "no Linux, no Node" framing — "Pretty sure Anthropic uses both Linux and node somewhere in their stack" (cap11235) — to which the creator responded that the point is no OS on the device itself. Beyond press coverage, organic developer testimonials remain scarce relative to the project's star count.
Bottom Line
MimiClaw is the ultimate minimalist statement: your AI assistant can run on $5 hardware drawing half a watt. It's not for everyone — flashing ESP32 firmware requires technical skill, API costs on the default Opus model burn fast, and development has visibly cooled since April 2026 — but for tinkerers who want the smallest, cheapest, most power-efficient option, MimiClaw remains remarkable.
Recommended for: Hardware tinkerers who want the cheapest, smallest, most power-efficient AI assistant possible.
Not recommended for: Users who want easy setup, multiple channels, or enterprise features.
Research by Ry Walker Research • methodology