Muse Lands on ESP32: Pocket Agent, Learning Crab Robot and a Pi-Sized Board

A personal agent on ESP32: Muse Gadgets

Meta has opened its Muse personal agent to small hardware. According to the independent MuseDirectory guide, the opening was announced on October 2, 2026 with two do-it-yourself routes: an ESP32 board or a RaspberryPi . On the ESP32 side everything runs through open-source firmware kept in the facebookincubator repository on GitHub ; the firmware is flashed onto the board, the board joins the home Wi-Fi network and pairs with a personal Muse account using a token taken from gadgets.muse.ai. Tokens start with mgst_ and can be revoked and reissued from the same page if they leak. Consumption above the free weekly allowance moves to paid plans.

The fastest route is letting a coding agent do the build. AGENTS.md files in every folder tell the tool how to install the toolchain, build for each board and flash it; Meta's own Muse Code arrives ready for this flow, and any agent that reads AGENTS.md can do the same job. The guide's advice is concrete: a speaker-equipped board for press-to-talk, an e-ink panel for a day display, the ESP32-C5 DevKitC-1 for learning. Two technical details matter: the cable must carry data, and the computer must run macOS or Linux; most charger cables carry power only and flashing fails silently. The firmware builds with version 6.0.1 of Espressif's ESP-IDF toolchain, and no other version is supported.

A pocket agent: Muse Charm

The Muse Charm shown on the Meta Connect stage revives the nineties Tamagotchi memory with modern internals. According to The Verge , citing Bloomberg, the device carries a two-inch OLED touchscreen, a fingerprint reader, cameras, speakers, microphones, a USB-C port and a built-in 5G modem that keeps it working without Wi-Fi; setup runs through the Muse app on iOS and Android. The plan is to ship in December in time for the holidays, pricing is rumored in the smartwatch range, and carrier fees for 5G remain open. Nearby Charms recognizing and interacting with each other is also on the table.

The pitch is personality. Muse gets a name, the cute on-screen avatar can be restyled, and the agent works with services such as Gmail and Drive, Shopify, OpenTable, Ticketmaster and PayPal, producing proactive suggestions once it can see the calendar and inbox; NewAtlas lists examples like ordering flowers before a trip and resolving clashing meetings by email. The same NewAtlas piece asks the reverse question: is separate hardware really needed for this job? The 199-dollar Rabbit R1 from 2024 is a reminder that dedicated agent gadgets stumbled in reviews yet still sold a hundred thousand units. Meta's answer is a speed claim: with no glasses worn, this palm device will be the fastest way to talk to Muse.

A crab that learns to walk: Jumper

The six-legged, 22-joint Jumper carries its crab-like body while using its front pair of legs as grippers. The project differs by teaching motion in simulation instead of coding it move by move: the KingKongRobotics repository on GitHub runs the MuJoCo Warp engine for GPU training and stock MuJoCo across CPU threads, exercising dozens of virtual robots at once through repeated falls and recoveries with the PPO method. Training commands ship ready-made; tripod gait, jump and dance tasks start with one command, and a trained policy can move to the on-board NPU through ONNX. The repository is also ready for coding assistants: AGENTS.md files explain what the project is and how training starts.

Honesty requires the caveats. The dance demo follows a prepared choreography; it is not a learned policy. Project docs state plainly that running trained models on the board and driving real motors with the new controller still await verification. Skin and scene tooling is only partly integrated. The direction still looks right: the next step in robotics is teaching motion in simulation rather than memorizing it. Jumper is one of the entry points bringing that threshold down to the home workshop.

A clock from counters and comparators

The week's recommended build is no Arduino shortcut; it is discrete logic the Eater way. Ben Eater's breadboard digital clock divides the 50 Hz mains signal down to 1 Hz with a comparator and advances time with counters; a four-part English series teaches digital electronics hands-on. The kit in Eater's shop costs $99.99 and takes a weekend or two; the box yields a 12- or 24-hour build with parts for 60 or 50 Hz mains. Buyers should pick the adapter for local voltage, 120-volt US or 240-volt EU. Subtitles make the English narration easy to follow.

A microcontroller in Pi 5 clothing

The week's value note comes from Waveshare . The ESP32-P4-Module-DEV-KIT is a credit-card-sized board, 85 by 56 mm, whose mounting holes and connector layout match the Raspberry Pi 5, so Pi cases and accessories fit. Inside sits a 400 MHz dual-core RISC-V ESP32-P4 with an AI instruction extension and single-precision FPU; wireless goes to an ESP32-C6 with Wi-Fi 6 and Bluetooth 5. Memory is 32 MB of PSRAM plus 16 MB of flash with a microSD slot on top; imaging brings MIPI camera and display links with 1080p30 H.264 encoding, plus Ethernet, four USB ports and a 40-pin header in Pi arrangement. The CNX Software review sums it up: Linux-board looks, microcontroller inside.

The warning label ships in the box: this is no Raspberry Pi. No Linux installs, no container stacks; in practice one program runs and development depends on Espressif's ESP-IDF framework, with Arduino and MicroPython trials still unstable. The trade is strength in real-time work: machine control and other critical processes that need instant hardware access suffer no operating-system latency. Thirty-two megabytes cover most small-server and light-application scenarios; this is not for buyers expecting an M.2-equipped Pi 5, it is for repetitive workshop jobs.

Memory prices set the budget

The appeal belongs to the same picture as the RaspberryPi price rises. Per the company's official notice, the LPDDR4 cost surge comes from AI infrastructure investment swallowing memory fab capacity; increases run $10 on 2 GB models, $15 on 4 GB, $30 on 8 GB and $60 on 16 GB. PCWorld reports the 16 GB Pi 5 moving from $120 to $220; consolation entries include a $83.75 3 GB Pi 4 while the $35 1 GB Pi 4 stays protected. Zero and third-generation boards escape the rises on older LPDDR2 stocks. The upshot: a microcontroller alternative for every job that needs no Linux is now a budget decision, not just a hobby.