Vibe CLI on Smart Pi One¶

Vibe CLI is Mistral AI's official Vibe command-line assistant, packaged by YUMI-LAB to run on the 32-bit ARM (Allwinner H3 / armv7l) Smart Pi One and Smart Pad — hardware most modern AI CLIs cannot run on without emulation.

Repository: github.com/Yumi-Lab/vibe-cli-smartpi

1. How it works¶

Vibe CLI is a thin client: the command-line tool runs locally on the board, but the AI inference happens remotely on Mistral's servers (model mistral-vibe-cli-latest). This keeps the heavy computation off the low-power Allwinner H3 — the board only handles the agent loop (tool calls, file edits, parsing), so there is no emulation and no version pinning. It installs as a native Python package via uv.

Because inference is remote, response latency is dominated by the Mistral API round-trip rather than the board's CPU.

2. Requirements¶

armv7l / 32-bit ARM CPU (Allwinner H3 or compatible Cortex-A7)
At least 1 GB RAM (swap protection via earlyoom)
A Debian-based Linux distribution with build essentials (tested on the Smart Pad — 4× Cortex-A7 @ 1.2 GHz, 1 GB RAM, Debian 13 armhf)
A Mistral API key — create one at console.mistral.ai

3. Installation¶

Run the one-line installer on your Smart Pi One as a normal user (not root):

curl -fsSL https://raw.githubusercontent.com/Yumi-Lab/vibe-cli-smartpi/main/install.sh | bash

The first install compiles a few native dependencies from source and takes ~15 minutes on the H3 (the build is bound to 2 CPU cores to keep temperatures in check). The installer adds ~/.local/bin to your PATH via ~/.bashrc and ~/.profile — open a new shell afterwards.

4. Authentication¶

Vibe CLI needs a Mistral API key (unlike Claude Code and Grok CLI, which sign in with a subscription account). Provide it in one of three ways:

# 1. Interactive setup wizard
vibe --setup

# 2. Headless — write it to the env file
mkdir -p ~/.vibe && echo 'MISTRAL_API_KEY=sk-...' >> ~/.vibe/.env

# 3. Per-shell export
export MISTRAL_API_KEY=sk-...

5. Usage¶

Command	Purpose
`vibe`	Full interactive TUI with prompts and session management
`vibe -p "your question"`	One-shot answer in programmatic mode
`vibe -p "your task" --yolo`	One-shot with auto-approval for all tool calls
`vibe --output streaming -p "…"`	Streaming, newline-delimited JSON output
`vibe -c` / `vibe --resume`	Continue or resume the previous session
`vibe --setup`	Configure your Mistral API key
`vibe --check-upgrade`	Check for and install updates

In agent mode, Vibe CLI can read and edit files and run commands on the board — useful for scripting the GPIO sensors, configuring services, or working through Smart Pi projects.

Example Vibe CLI session on a Smart Pi One Example session (illustration).

6. Notes¶

Thin client: your code and prompts are sent to Mistral's servers for inference. No large model runs on the board.
Thermals: the one-time install compile is the hot part (~87 °C peak, 2-core bound); at rest the board sits around 68–70 °C.
Cold start: the Python runtime adds a few seconds of startup latency per invocation (e.g. a version check takes ~7 s).
Licensing: the installer scripts are MIT (YUMI-LAB); Mistral Vibe itself remains subject to Mistral AI's terms and is installed from the official distribution at runtime.