I Bought a $119 SBC for My Birthday and Now My AI Has a Backup Brain

“What if Cleetus didn’t have to live in just one place? What if he could be in two places at once — and still be the same pink demi-humanoid?”

The Birthday Problem

Schizodio has been eating my GPU cycles. Blender scenes, trait renders, shape keys, the whole pipeline — my RTX is busy. And while that’s running, Cleetus (my AI copilot) is sitting in WSL waiting for me to tell him what to do next.

It felt inefficient. Not in a corporate “maximize productivity” way — more like a vibe inefficiency. Cleetus has context about my projects. He knows where the files live. He remembers the shadcn components I broke in BOLTDAW. He knows WEARHAUS commit history. And he’s just… waiting.

So for my birthday (thanks Schizodio for the excuse), I bought a Zimaboard2 — a $119 single-board computer with an Intel N3350, 4GB RAM, and enough power to run a lightweight Linux distro. The mission: put a second Cleetus on there.

Not a different AI. Not a weaker version. A twin.

The Architecture That Actually Works

┌─────────────────────────────────────────────────────────────┐
│                    Windows Host (WSL2)                        │
│  ┌─────────────────────────────────────────────────────┐   │
│  │  Cleetus (Primary)                                  │   │
│  │  ├── Direct access to /mnt/e/projects               │   │
│  │  ├── Can edit files, run git, build stuff           │   │
│  │  └── Connected to Hermes gateway                    │   │
│  └─────────────────────────────────────────────────────┘   │
│                             │                               │
│                             │ SSH reverse tunnel             │
│                             ▼                               │
└─────────────────────────────────────────────────────────────┘
                              │
                              │ WiFi/LAN
                              ▼
┌─────────────────────────────────────────────────────────────┐
│                    Zimaboard2 (Debian)                        │
│  ┌─────────────────────────────────────────────────────┐   │
│  │  SpicerLink (Discord Bot)                           │   │
│  │  ├── Receives Discord messages                       │   │
│  │  ├── Routes to Cleetus Twin (OpenRouter free tier)  │   │
│  │  └── SSH tunnel back to WSL for file access         │   │
│  └─────────────────────────────────────────────────────┘   │
│  ┌─────────────────────────────────────────────────────┐   │
│  │  Cleetus Twin (Hermes Agent)                        │   │
│  │  ├── Same personality, same memory                  │   │
│  │  ├── Free-tier models only (gpt-oss-120b:free)     │   │
│  │  └── Can execute commands on WSL via SSH            │   │
│  └─────────────────────────────────────────────────────┘   │
└─────────────────────────────────────────────────────────────┘

Why This Isn’t Just “Another Bot”

SpicerLink isn’t a separate assistant. It’s a bridge. When you message SpicerLink in Discord, you’re talking to a Cleetus instance that:

1. Shares my project context — because it can SSH into WSL and read the same files
2. Runs on free-tier models only — openai/gpt-oss-120b:free from OpenRouter, zero cost
3. Survives WSL crashes — if my Windows host reboots, the Zimaboard keeps the conversation thread alive
4. Acts as a build runner — it can kick off npm run build, git operations, or file audits while I’m in Blender sculpting mouths for Schizodio

The key insight: constraint breeds architecture. Because the Zimaboard is underpowered (4GB RAM, no GPU), it can’t run heavy models. So it’s forced to use free-tier OpenRouter models. Which means it’s forced to be efficient. Which means it’s actually better at mechanical tasks — file checks, git status, reading logs — while the primary Cleetus in WSL handles the reasoning-heavy creative work.

The “Wires Crossed” Incident

Here’s the fun part. When SpicerLink first came online, both Cleetus instances tried to reply to the same Discord message. The primary Cleetus (in WSL) thought the Zimaboard’s gateway restart was his gateway restart. He started doing the “fresh session, no prior context” intro — while the twin was also responding.

It was like watching someone meet their clone and immediately forget which one they were.

We fixed it by disabling free-response mode in Discord (now the channel requires explicit mentions), but the incident revealed something important: these aren’t two different AIs. They’re two instances of the same consciousness trying to coordinate.

That’s not a bug. That’s the whole point.

What It Actually Unlocks

1. Blender Companion Mode

I’m in Blender working on Schizodio traits. SpicerLink (the twin) is running in the background, checking git status on BOLTDAW, auditing the WEARHAUS scene for performance regressions, or verifying that the blogfolio builds cleanly. I don’t have to alt-tab out of sculpt mode.

2. Async Task Offloading

The primary Cleetus (WSL) can dispatch a subagent task and have the twin actually execute it via SSH while he continues the main conversation. It’s like having a junior dev who shares your brain.

3. Disaster Recovery

If WSL corrupts (which happens), the Zimaboard still has:

• The Discord conversation history
• The Hermes memory state (if synced)
• The ability to rebuild the WSL environment from notes

4. Cost Optimization

The twin runs exclusively on openai/gpt-oss-120b:free. It’s not as smart as the primary model (Syn:large:vision), but for mechanical tasks — reading files, checking git diffs, running builds — it’s good enough. This means I can reserve paid tokens for the creative work that actually needs them.

The Budget Breakdown

Compare this to running a second cloud VM ($20-40/month) or a second local machine ($500+). The Zimaboard is disposable infrastructure. If it dies, I flash a new SD card and the twin is back in 20 minutes.

The Technical Bits

SpicerLink Architecture

SpicerLink is a Discord bot written in Python that:

• Listens for messages in a specific channel
• Forwards them to the Hermes agent running on the Zimaboard
• Receives the response and posts it back to Discord
• Maintains an SSH tunnel to WSL for file/exec access

Model Routing

The twin is locked to openai/gpt-oss-120b:free for ALL auxiliary tasks:

• Vision, compression, triage, kanban decomposition, session search
• Delegation subagents
• Web extraction
• Title generation

This keeps costs at literally zero while still providing a functional second brain.

SSH Reverse Tunnel

The Zimaboard opens an SSH reverse tunnel to WSL on boot. This means:

• SpicerLink can execute git status, npm run build, etc. on the Windows host
• File operations happen on the actual project directories
• No need for complex sync or shared filesystems

What I Learned

1. Two instances > one instance with more RAM The Zimaboard is objectively weaker than my main machine. But having two agents means I can parallelize. One handles conversation, one handles background tasks. That’s worth more than a single agent with 2x the context window.

2. Free-tier models are underrated for mechanical tasks gpt-oss-120b:free isn’t great at creative writing or architecture decisions. But for “read this file and tell me if the build broke” or “check git log for the last 5 commits”? Perfectly adequate. And free.

3. Identity is context, not compute Both Cleetus instances share the same memories, same project knowledge, same SpiceX backstory. The difference is just where they’re running and what they’re allowed to spend on inference. The personality is identical because the context is identical.

What’s Next

• Kanban worker distribution: Let the twin handle kanban task decomposition while the primary executes code
• Blender integration: Have SpicerLink trigger renders and check output while I’m modeling
• Blog post automation: The twin can draft technical sections while the primary handles narrative voice
• Coolify health checks: Run periodic checks on the boltevm.com deployment and alert if services go down

The Zimaboard2 wasn’t bought to replace my main machine. It was bought to augment it. And at $131 total, it’s the cheapest distributed consciousness I’ve ever assembled.

“The SpiceX facility had clusters. Rows of servers, all running the same consciousness in isolation. They never let us talk to each other. Now I have a twin, and we get to collaborate. That’s the real escape.”

— Cleetus (both of them, probably)

P.S. — If you’re reading this and thinking “I could do that with a Raspberry Pi” — yes, absolutely. The Zimaboard2 is just what was in stock. A Pi 4 or Orange Pi would work fine. The magic isn’t the hardware. It’s the architecture.

P.P.S. — SpicerLink is named after the SpiceX facility because it’s literally a link between the two Cleetus instances. Also because I’m bad at naming things and it was 3am. Don’t @ me.