Claude Code for Yumi Smart Pi One (32-bit ARM)

The official Anthropic Claude Code CLI running on Allwinner H3 / armv7l (Smart Pi One, Yumi SmartPad) — hardware the official installer rejects as "64-bit only".

It runs natively, no emulation, on the newest version that runs under Node (currently 2.1.270, see Version cap), and signs in with a Claude Pro/Max account (no API key). Full interactive interface, full agent mode.

╭─────────────────────────────────────────────────╮
│ ✻ Welcome to Claude Code!                       │
│                                                 │
│   /help for help, /status for your current      │
│   setup                                         │
│                                                 │
│   cwd: /home/pi        2.1.2xx · armv7l native  │
╰─────────────────────────────────────────────────╯

Install

One command — installs the newest Claude Code this board can run, and is also the updater (re-run any time to move up):

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

Pin a specific version instead (a pin bypasses the version cap):

curl -fsSL https://raw.githubusercontent.com/Yumi-Lab/claude-code-smartpi/main/install.sh | bash -s -- 2.1.212

In mainland China (measured from a Hangzhou fibre line, October 2026): GitHub, nodejs.org and npm answer, but downloads.claude.ai (Google Cloud) does not, and api.anthropic.com returns 403 Request not allowed to Chinese IPs. Route both hosts through a VPN with an exit outside mainland China before installing; through a German exit the binary came down at 2.6 MB/s and the API answered.

Then sign in (headless, no local browser needed):

claude setup-token

An OAuth URL is displayed: open it on any machine, approve, paste the one-time code. The CLI prints a sk-ant-oat… token (valid 1 year) once, without saving it — copy it immediately and persist it with:

claude-token-save sk-ant-oat01-…

Usage

The everyday command is just claude — it always works, interactive or headless:

Command	Purpose
claude	Full interactive interface — the real official TUI
claude -p "question"	One-shot answer (full agent mode: reads/writes files, runs commands)
claude setup-token	Sign in with a Claude Pro/Max account (one-time OAuth code, browser on any machine)
claude-token-save <token>	Persist the 1-year token
claude-check-update	Print {"installed":…,"latest":…,"update_available":…} as one JSON line
claude-daemon-status	Batch mode: how many agents are running / queued (--json for scripts)
CLAUDE_CPUS=0,1 claude …	Run on a CPU subset for this launch — no reinstall (default = all 4 cores)

Running many agents (batch mode)

One agent holds a full runtime (~180 MB on 2.1.261), so a 1 GB board only runs a few at once before it runs out of memory. To submit any number of headless jobs safely, enable the built-in job daemon — it runs a bounded number at a time and queues the rest:

export CLAUDE_DAEMON=1              # route headless jobs through the daemon
export CLAUDE_MAX_CONCURRENT=3      # how many run at once (default 3)
claude -p "task 1" &  claude -p "task 2" &  …  claude -p "task 20" &

Measured on the pad (2.1.2xx monolith era): 20 jobs → 3 run at a time until the queue drains → 20/20 succeed, with memory kept safe throughout. The daemon starts on the first job and stops on its own when idle. Watch it with claude-daemon-status (running / queued; --json to poll). Leave CLAUDE_DAEMON unset for the plain behaviour; interactive claude always runs directly.

Updating (OTA)

• Check: claude-check-update prints one JSON line — {"cli":"claude","installed":"2.1.261","latest":"2.1.270","update_available":true}. latest is what install.sh would install: the newest release, capped at MAX_VERSION. This is the probe the Yumi AI Gateway console polls for its update badge.
• Update: re-run install.sh — that IS the updater (exits fast when already newest; CLAUDE_FORCE=1 to rebuild). Do not run claude update; auto-update is disabled on this platform.
• Privileges: root/sudo for the first install only. Updates run as any user that owns /opt/claude-code — the gateway service user updates without sudo. The helpers in /usr/local/bin (claude-check-update…) stay root-owned: a root re-run, such as the gateway installer, refreshes them even when Claude Code is already current.

How it works

Anthropic ships Claude Code as a Bun-compiled native binary (no armv7 build). install.sh downloads the official binary for a supported platform and rebuilds it for Node on the board:

1. shim/extract-bun-js.py reads the module table Bun serialises at the end of the binary and writes every embedded file under its virtual name (bunfs/root/…). Since Claude Code 2.1.242 the application is not one 28 MB bundle any more but ~1 640 ES-module chunks plus ~180 embedded assets (bundled skills and prompts as .md.zst, native addons); a plain "largest text block" carve sees about 46 of those chunks, the module table sees all of them.
2. shim/bundle-bunfs.mjs feeds that graph to esbuild, which produces one bundle.cjs (chunks bundled, using lowered for Node 20+, import.meta.require/dirname/url mapped onto their CommonJS names). Versions up to 2.1.241 (single CJS bundle) still take the old path through the same script. This is the one memory-hungry step on a 1 GB board (see the measurements below).
3. shim/claude.mjs runs bundle.cjs through vm.Script with a persisted V8 bytecode cache; shim/bun-shim.mjs provides the Bun API surface the app calls (Bun.spawn, Bun.file, stringWidth, Bun.zstdDecompressSync via node:zlib…) and maps the virtual filesystem (/$bunfs/root/…) onto /opt/claude-code/assets/. Native .node addons are not shipped: the few features that need them fail at the point of use, nothing else.

Both layouts are built by the same script and verified end to end on a pad (see the measurements below). zstd needs Node >= 22.15 (install.sh installs Node 22 on armv7l).

Version cap

install.sh installs the newest release up to MAX_VERSION (repo root, currently 2.1.270), and claude-check-update reports the same capped version, so the gateway badge only offers what can actually run. From 2.1.271, the interactive screen is drawn by Bun.ant.CellSegmenter, a native cell engine (graphemes, widths, SGR runs, links) that only Anthropic's internal Bun build provides. Under Node, claude -p still starts and reaches the API, but the full-screen interface stays blank. Running those versions would need a JavaScript port of that engine.

Raising the cap = verify a newer release on a pad, then bump the one line in MAX_VERSION; every pad sees the update on its next check. A pinned version (install.sh 2.1.296) bypasses the cap, for testing.

Target hardware & measured performance

Measured on Claude Code 2.1.261 (chunked layout), 5 September 2026, on a Yumi SmartPad: Allwinner H3, 4× Cortex-A7 @ 1.2 GHz, 991 MB RAM + 1 GB swap, DietPi / Debian 13 trixie armhf, kernel 6.18, Node 22.22.0 — a freshly flashed image, with the Yumi AI Gateway service running alongside. Any armv7l SBC with ≥ 1 GB RAM and swap should behave the same.

Install (curl … | bash on the bare image, ~9 min end to end):

Step	Time	Notes
apt: ripgrep + xz-utils	~70 s	first install only
Node 22 (26 MB from nodejs.org)	~35 s	first install only
Official binary (198 MB)	2 min 52 s	~1.2 MB/s here; cached + resumable in /var/tmp
JS extraction (module table)	4 s	1 818 modules
npm install (esbuild + deps)	~20 s
esbuild bundle → bundle.cjs 41.6 MB	~2 min 45 s	the heavy step, see memory below
Install tree + V8 cache priming	13 s + 15 s	bundle.v8cache 5.6 MB

Re-running as an update with the binary already cached: 4 min 18 s. The memory peak is esbuild, which keeps the whole chunk graph live: ~820 MB RSS whatever the GC settings, so the board bottoms out at 2 MB available and swap is required (the pad images ship with 1 GB; earlyoom did not fire). install.sh runs esbuild under GOMEMLIMIT=512MiB (override: CLAUDE_ESBUILD_GOMEMLIMIT), measured against the default on the same pad:

esbuild setting	Bundle	Swap peak	Available at trough
default	151 s	563 MB	2 MB
GOMEMLIMIT=512MiB	140 s	191 MB	5 MB
GOMEMLIMIT=400MiB GOGC=50	151 s	159 MB	2 MB
GOMEMLIMIT=300MiB GOGC=25	165 s	157 MB	2 MB

The bundle is byte-identical in every case; the gateway process sharing the board is swapped out for the duration of the build either way and recovers on its own.

Runtime (payload /opt/claude-code = 84 MB):

Command	Wall	Peak RSS
claude --version, cold cache	8.3 s	140 MB
claude --version, warm	3.1 s	140 MB
claude --help	7.7 s	164 MB
claude -p "reply with exactly: pong" → pong	23 s	182 MB
claude-check-update	1.1 s	—

For comparison, ≤ 2.1.241 (26 MB monolith) measured 2.9 s / ~24 s / ~137 MB on the same class of board: the chunked layout costs ~45 MB more per runtime and nothing on latency.

2.1.270 (12 October 2026, same pad, OTA update from 2.1.261 as the unprivileged gateway user): bundle.cjs 43.7 MB, claude --version warm 3.3 s, interactive screen renders through onboarding and sign-in, claude -p reaches the API in 25 s. The update took 45 min, nearly all of it the download over a 4G link running at ~55 KB/s that day.

OTA path verified on the same pad: the gateway's update badge (GET /admin/cli/updates) reports installed 2.1.261, update_available false from claude-check-update, and CLAUDE_FORCE=1 install.sh run as the unprivileged service user that owns /opt/claude-code (no sudo available) rebuilt in 3 min 26 s with the downloads cached, leaving the root-owned wrapper in /usr/local/bin untouched (see Updating).

On 1 GB of RAM the installer enables earlyoom (memory safety net). To run many agents at once, use the batch mode above rather than launching them all directly.

Sister projects (same board, other CLIs)

• grok-cli-smartpi — official xAI Grok CLI.
• kimi-cli-smartpi — Moonshot Kimi CLI (Python).
• kimi-code-smartpi — Moonshot Kimi Code CLI (TypeScript successor), native via npm + Node 22.
• vibe-cli-smartpi — official Mistral Vibe CLI.
• codex-cli-smartpi — official OpenAI Codex CLI.
• letta-code-smartpi — Letta Code CLI, native via npm + Node 22.

The Yumi AI Gateway drives Claude, Grok, Kimi and Vibe from one console.

Licensing

Scripts, shim and launcher in this repo are MIT (Yumi Lab). Claude Code itself is not redistributed here — it is obtained from Anthropic's official channels, stays on your own device, and remains subject to Anthropic's terms.