Skip to main content

Nimbus

Nimbus at an angle, showing the speaker grille, power switch, and ports

An open-firmware AI assistant for your desk, on a commodity ESP32-S3 board.

Orchestrator mode: reach your always-on Nimbus over Telegram, hold-to-talk on the device, or use the local web UI. Nimbus remembers across conversations, runs routines, uses tools and connects to your favorite MCPs.

Notifier mode: each coding session takes a segment on the LED ring, communicates over encrypted Bluetooth, so a glance shows which agents have finished and what needs you.

Memories and data stay on the device, not on a cloud account. Only the model calls leave, to the provider you pick: a hosted lab, or an LLM you host yourself. Bring your own key (Mistral, OpenAI, Anthropic, or OpenAI-compatible). Source-available, noncommercial.

Get started Flash from your browser
License - source-available, noncommercial

Code is under the PolyForm Noncommercial License 1.0.0; documentation and hardware are under CC BY-NC-SA 4.0. Build it, modify it, and share your fork, for any noncommercial use. Details in the repository's LICENSE, LICENSE-docs, and NOTICE.

Two builds, one firmware​

Same firmware, same features on both builds. They differ in assembly effort, and in whether the status ring is physical or rendered on screen.

The hand-built Classic Nimbus in its 3D-printed case

Hand-built

Classic Nimbus

An ESP32-S3-DevKitC-1 on a custom carrier PCB in a 3D-printed case, with a discrete WS2812B ring and a 2.8-inch touchscreen. Soldering and assembly required.

  • ESP32-S3-DevKitC-1 (N16R8) + carrier PCB
  • WS2812B LED ring
  • 2.8" ILI9341 240x320, resistive touch (XPT2046)
  • Mic, speaker, microSD, optional 2S battery
  • Around $60 to $75 in parts

Build guide →

Render of Nimbus Light in its printed case on the desk stand

All-in-one

Nimbus Light (Freenove CYD)

A single Freenove ESP32-S3 module: display, capacitive touch, microSD, mic, and speaker on one board. No wiring, no soldering. The status arcs render on the panel.

  • Freenove ESP32-S3 Display (FNK0104B)
  • 2.8" ILI9341 240x320, capacitive touch (FT6336U)
  • Status ring rendered on-screen
  • ES8311 mic + speaker, microSD, optional 1S battery
  • Around $20 to $25, one board

All-in-one guide →

Every panel is a 2.8" ILI9341 (240x320) on an ESP32-S3 N16R8 (16 MB flash, 8 MB PSRAM); the Freenove also comes in 3.5" and 4.0" sizes. Pinouts and wiring are in the hardware reference; the parts list is in the bill of materials.

One device, two modes​

Pick a mode at flash time; switch later from the settings menu. A switch restarts the device.

Notifier​

A broker on your computer speaks the nimbus-notify protocol over an encrypted, paired Bluetooth link. Each coding session maps to a ring segment; its color and animation encode the state: running, waiting for input, done, errored. On Nimbus Light the segments render on screen.

Notifier quick start

Orchestrator​

A self-hosted agent backed by a hosted LLM and your own key. Reach it over Telegram, hold-to-talk on the device, or the local web UI. Long-term memory on the SD card, tools mid-turn, background sub-agents, scheduled routines, and signed over-the-air updates.

What you need

What stays on the device​

Memory, routines, and tool actions all happen on the device you own. Only the model call ever leaves, and only when you point it at a hosted provider.

Self-hosted model

Your home network
Nimbus memory + routines
↓ model call
Your computer self-hosted LLM, MCP tools

Host the model yourself and even the model call stays on your network. Nothing crosses out.

Hosted provider

Your device
Nimbus memory + routines + connector tools
↓ model call
AI lab Mistral, OpenAI, or Anthropic

Point at a hosted provider and only the model call leaves. Memory, routines, and actions stay on the device.

Cumulo Nimbus - the first-party service

Nimbus runs entirely on your own keys and network. Cumulo Nimbus is an optional first-party service the firmware integrates with directly. Off by default.

  • One key across providers. A Cumulo key draws on one prepaid balance routing to Mistral, OpenAI, and Anthropic. Paste it under Assistant → Models, or use it from your own code as an OpenAI- or Anthropic-compatible base URL.
  • Remote access without port forwarding. Cloud access opens one outbound secure tunnel to reach the device's web UI from anywhere. Pair it with a Cloud link code shown on the device.

Both opt-in, both verify ownership before anything reaches the device, and neither replaces the local web UI.

Where to go​

note

These pages are a published view of the firmware repo's docs/ tree. The canonical source is the Markdown in the repository; see the docs map.