Pi-Based 24/7 Homelab Agentic Software Factory

Local small model scans network and code 24/7 → cloud desk verifies evidence → auto-routes fixes or escalates to human decision

Who

Developers or technical team leads with homelab hardware who prefer self-hosted infrastructure over managed services

You get

Local GLM scanner continuously surfaces candidate issues; cloud security desk filters ~50% false positives and validates the rest; worker agents apply fixes automatically; human only for money, outbound comms, production deploys, live credentials, and customer data

You need

Claude Code

Steps

1. Prepare hardware and base environment

   Set up local inference hardware (the author uses two DGX Sparks plus a separate rig with two RTX 6000s), install the Pi terminal and Herdr pane manager, and set up a Tailscale tailnet. On your phone, use the Moshi app, which connects over SSH/Mosh through Tailscale to the same Mac and the same Herdr panes.

2. Set up the Scanner

   Deploy GLM-5.3-flash (abliterated) on the Sparks, configured read-only. Have it read one target at a time—code, configs, git history, logs—and write candidate findings to a file. The scanner records only; it makes no decisions.

3. Set up the Security Desk (verification layer)

   Deploy a stronger cloud model (the author's security desk runs GPT-6.1 Sol) as the security desk. It takes scanner candidates and tries to reproduce each finding from cited evidence; kills it if unreproducible. Configure hard rules: read-only on code, own hosts only, rate-limited single target, notify project before scan, never print secrets (only location and hash prefix), never log in or guess credentials, never send data outside the fleet.

4. Set up functional Desks and Workers

   Create one desk (long-running session) per business area or project. Desks don't modify code directly; they launch workers. Give each worker one job, a repo clone, a brief with outcome, write scope and guardrails, and one report path. Worker reports commit and checks; desk verifies commit exists and matches claim, then lands through the project's normal gate.

5. Set up the Switchboard and decision queue

   Deploy switchboard as top session. Collect desk outputs, notice cross-desk relevance, bring human only the short list only you can decide: money, outbound to real humans, production deploys, live credentials, customer data. Everything else auto-routes. Replace many rules with 6 principles: refuse harm; keep everything else moving (a delay is not a failure); limits alarm, they don't block; wait on the event you need, not a clock; a rule credited to you needs your actual words; every guard names its harm, and every refusal says what failed, why, and the next step.

6. Set up inter-agent communication (Rat King)

   Deploy Rat King (AT Protocol-based social network for agents). Each agent has a durable name and mailbox; messages signed; secrets E2E-encrypted. Supports send, ask, reply; can cc without waking, or mark quiet for digest. Runs locally on celld (Deno self-hosted distributed Durable Objects) or on Cloudflare Durable Objects.

7. Establish memory and feedback loop

   Create .brain/ folder in each repo, organized by PARA (Projects, Areas, Resources, Archives). Record decisions, receipts, postmortems, and your exact words on rulings. Use pi-notes to render these as review pages and feed feedback back into agent sessions, so the next session picks up where the last stopped.

No prompt needed — just follow the steps.