What I run when I write code with agents, laid out on a Raspberry Pi because every part has an obvious slot there. Scroll to take it apart.
Parts
- 01 rex terminal · multiplexer · the PCB
Terminal and multiplexer in one: sessions that outlive the window, a tab per agent, scriptable input, and everything below mounts here.
- 02 pi coding agent · the SoC
Small core, any model, and every part of it is a TypeScript extension I can read rather than a product I configure.
- 03 fable 5.1 · high oracle · the microSD
What the board boots from: it plans, reviews and decides what gets fixed, at high effort because a wrong call here costs the most.
- 04 opus 5.5 · med workers · the RAM
One per slice, several at once, briefed and then checked by the oracle; medium effort is plenty once the task is specified.
- 05 gpt 6 luna · low scout · the wireless
Sent out first to read, grep, fetch and summarise; cheap enough to send often, low effort because it only reports back.
- 06 agent-skills workflow · addy osmani · the GPIO header
Planning, TDD, review loops and shipping as skills the agent loads when they apply, one pack per session.
- 07 BrowserSkill chrome devtools · the HDMI
The display output: a real browser in the loop so the agent can verify, dogfood and profile what it just built.
- 08 mcp servers context7 · notion · slack · the USB
Peripherals: docs, notes and conversations plugged in per task and unplugged when done.
- 09 context-mode context sandbox · the ethernet
Big outputs run in a sandbox and only the answer comes back, so logs and test runs stop eating the context window.
- 10 grove worktrees · the camera port
One worktree per task across several repos, so parallel agents never edit the same checkout.
Drawn in code with Zdog, to the Raspberry Pi 4 footprint. The setup changes often; the date above is the last time this page matched it.