A decommissioned Dell PowerEdge R640, reinstalled with Ubuntu Server 26.04.1 LTS and set up as a shared AI workspace for Firewheel's mentors and students. Chad installs it on site.
No license, no subscription, no prolonged maintenance.
Firewheel has no sysadmin. Install once, walk away, still working in three years. If an option needs someone to tend it, it is the wrong option here, even when it is technically better.
Each layer runs on the one below it. Solid boxes ran end to end in the rehearsal on a real Ubuntu 26.04.1 install. Dashed boxes are designed but not fully built; they are covered in the AI brain section below.
add-mentor.sh/opt/ms-playwrightresolute repo, Compose v5.5npxinstall/stack.json, rehearsal run 2
Only two doors face the LAN. Everything else listens on the box itself. Nothing faces the internet: no port forwarding, no dynamic DNS.
Launched in 2017. Dell's 14th-generation 1U server: dual-socket Intel Xeon Scalable (Skylake or Cascade Lake), up to 24 DIMMs, iDRAC9. For this workload it is comfortably overqualified. Claude's compute runs at Anthropic, so the box needs a shell, RAM for editors and containers, and a network.
| Verified fact | Value |
|---|---|
| Form factor | 1U, 42.8 mm high, 482 mm wide at the rack ears |
| Depth | 769.66 mm body · 808.51 mm to PSU handles |
| CPU sockets | 2 × Intel Xeon Scalable |
| Memory slots | 24 DIMMs |
| Management | iDRAC9, Express or Enterprise |
| Drivers | All in the Linux kernel: megaraid_sas, igb / bnx2x, ast |
| Dell's Ubuntu matrix | Validated up to 22.04. Dell stopped testing 14G there, which applies to 24.04 and 26.04 alike |
| GPU | Up to 3 single-width cards (T4 class). Not used: no local AI models |
| End of life | No Dell date published. Firmware stays free without a contract |
| Power supply | Input voltage |
|---|---|
| 495 W Platinum | 100-240 V, autoranging |
| 750 W Platinum | 100-240 V, autoranging |
| 750 W Titanium | 200-240 V only. Will not run on a standard 120 V outlet |
| 1100 W Platinum | 100-240 V, autoranging |
| 1600 W Platinum | 100-240 V, autoranging |
| Running cost, estimated | |
|---|---|
| Real draw, 24/7 | 150-350 W |
| Electricity per year | about $250-450 |
| Replace with the real number from the iDRAC power readout once it is running. | |
| Noise, by configuration | CPU | Idle | Working | Scale 0-50 dBA |
|---|---|---|---|---|
| Minimum | 1 × Bronze 3104, 85 W | 35 dBA | 37 dBA | |
| Typical | 2 × Gold 5120, 105 W | 36 dBA | 39 dBA | |
| Feature rich | 2 × Gold 6142, 150 W | 39 dBA | 46 dBA | |
| Dell, 14G Acoustical Performance, 2018. Blue bar idle, amber bar working. Every power-on runs the fans at 100% for about 30 seconds. Dell calls the minimum build "quiet enough for typical office environment." | ||||
The exact configuration of this unit is not known yet. These get read off the machine itself. The first three can each stop the install if they go the wrong way.
Read the PSU label. 750 W Titanium needs a 208/240 V circuit. Also check that both supplies are present.
Are drives included, with caddies? Empty bays need blanks, or airflow suffers and the fans get louder.
4 × 10 GbE SFP+ and 2 × 25 GbE have no RJ45. Fix is a ~$30 PCIe gigabit card.
Model, socket count, DIMM count. CPU sets power, heat and noise.
H730P / H740P is ideal. H330 has no cache and weak RAID 5/6, but RAID 1 or 10 is fine on it.
If present, the OS goes on its M.2 pair and every front bay holds data. Reinstalls never touch the data.
No BIOS setup password and no iDRAC System Lockdown. Lockdown plus lost iDRAC credentials is the one true brick.
Rails and C13/C14 power cords, included or not.
A rack needs about a metre of depth. Without one, a shelf that carries the full depth, not two corners.
Air-conditioned. A sealed closet in a Chandler summer will shut it down.
A 120 V circuit that can carry it. A UPS is worth considering.
A wired drop, ideally two so iDRAC gets its own.
Everything after the Ubuntu installer is two scripts, both safe to re-run. Rehearsed from a blank disk to a clean unattended reboot.
ubuntu-26.04.1-live-server-amd64.iso and check it against SHA256SUMSclaude-toolkit and ai-os-brain. Bundles, never folder copies: Windows line endings make every file look modified on Linuxubuntu-lv to the max. The default gives root half the disk, capped at 100 GB. Delete the old size before typing the new onefirewheel, user fwadmin. Skip Ubuntu Pro. Tick "Install OpenSSH server", which is off by default. No snapsssh-copy-id fwadmin@<box-ip> git clone /media/usb/firewheel-server.bundle /srv/firewheel/firewheel-server
firewheel-install.sh inside tmuxIt downloads several GB and refuses to start outside tmux, because a dropped SSH session killed a build in the rehearsal.
sudo bash install/add-mentor.sh <username> <their-key.pub>
Makes a key-only account, installs their own Claude Code and connects Crawl4AI and Firecrawl. They sign in to Claude the first time they run claude.
Reboot once without touching anything, then confirm everything came back on its own with sudo python3 install/access-probe.py --print. Fill in and tape up the card on the closet door.
Two full installs in VirtualBox from the exact ISO, before the hardware arrives. Anything Dell-specific (PERC, iDRAC, BIOS, PSU, NIC) waits for the real box.
The 9 yellow checks are the wiki brains and repos waiting for a Firewheel GitHub org. Two test mentors were added, each with a self-updating Claude Code and both web tools connected.
Tested: Docker publishes for it, and Node 22 is in Ubuntu's own repos. Four more years of free updates than 24.04's window started with.
Both attach a token to the machine that can quietly lapse.
Kernel fixes need a reboot without Livepatch. Nobody is logged in at 4 AM. This one setting is most of the no-maintenance story.
The PERC handles disks with no software to understand. Boring is the point.
Email alerting needs an SMTP relay and a mailbox someone maintains. The drive light and a phone number on the wall will still work in three years.
Nothing expires, nothing to reset. The script skips this step if your key is not on the box yet, so you cannot lock yourself out.
Docker ports ignore the firewall. Binding to localhost keeps the vector store and search off the Wi-Fi.
A global npm install is owned by root and never updates. Per-user copies update themselves, and no API key sits in a shared file.
A re-run in two years rebuilds the exact set that was tested, not whatever upstream shipped that week.
Kubernetes, Proxmox, VMware, Swarm, a monitoring stack, configuration management. Each one is something a person has to understand later.
Firewheel GitHub organisation. The six repos have no remote on purpose: brains on a personal account would be a loan, not a donation. Fredi or Glen creates the org with Chad as admin. Until then there is no off-box backup and no cross-brain querying.
What the four project brains are for. They are numbered, not named, until the server's use is discussed with Firewheel.
How mentors see the BRAIN dashboard. Today it is an SSH tunnel. Opening it to the LAN needs a firewall rule, and the page has no login.
Claude seats. Firewheel qualifies for Anthropic's nonprofit rate. The form needs a Firewheel staff email and title, so Firewheel submits it.
The 04:00 reboot actually firing, a staff member logging in to Cockpit, and RAG ingest, which is not built yet.
Two kinds of memory, for two kinds of question. The AI brain is a small set of pages Claude writes and keeps up to date, so it can answer "where are we and why" in seconds. The data layer holds every manual, rulebook and web page in full, so Claude can find the exact passage. Claude checks the brain first and the data layer second.
raw/ and wiki/ hold anything irreplaceable, and both are plain files in git.| AI brain | Data layer | |
|---|---|---|
| Answers | "Where is this project, what did we decide, and why?" | "What exactly does the manual say about X?" |
| Holds | A few dozen to a few hundred short pages per project | Every source document in full, thousands of passages |
| Who writes it | Claude, from what people put in raw/ | The ingest script, with no judgement involved |
| How it finds things | Reads the index, follows links between pages | Searches by meaning, not exact words |
| Where it lives | /srv/firewheel/firewheel-*-wiki | Qdrant in Docker on 127.0.0.1:6333, plus catalog.db |
| If it is lost | Restore from git. Pushed off the box once the Firewheel org exists | Re-run ingest over raw/. Nothing is lost |
Someone drops a file in a brain's raw/ folder. Claude reads it, says what it learned, writes a source page, updates every topic page it touches, then updates the index and log. The same file also goes through the data layer's chunk-and-embed step.
Claude opens the brain's index first and follows only the pages it points to. For detail it searches Qdrant. Every answer names the source file behind it. A good answer can be saved as a new page.
On request, Claude reports broken links, orphan pages, claims with no source and contradictions between sources. It reports and does not auto-fix, so a person sees each change.
A small local model (all-MiniLM-L6-v2 on ONNX Runtime) turns text into vectors on the CPU. It needs no API key or account, so there is nothing to renew. The same setup answers searches in about 0.1 s on Chad's own video-transcript corpus. A hosted embedding service would mean a key and a bill.
The record is the files in raw/. If the vector store breaks, delete it and run ingest again. Nobody has to back it up or repair it.
Inventory questions get an exact answer from one SQLite query, not a guess from a folder listing.
No database to administer. Students can read the pages in any editor. Git history shows who changed what.
Once Firewheel's GitHub org exists, each brain pushes to it nightly with no one involved. Until then, the brains live only on the server.
| Piece | State |
|---|---|
| Master brain and four project brains | Built Scaffolded and committed, local git, no remote |
| Qdrant vector store | Built Installed by the script, running on 127.0.0.1, empty |
| Chunk-and-embed ingest script | To build Same pattern as Chad's ICT corpus, with the local CPU model |
| catalog.db | To build Same pattern as the ICT corpus catalog |
| Search tool for Claude Code (Qdrant) | To build An MCP server per mentor, like Crawl4AI is wired in today |
| Cross-brain search | Proposed Local tool reading the brain folders. Chad's own version reads GitHub and only lets Chad in, so it does not carry over |
| Nightly push of the brains | Blocked Needs the Firewheel GitHub org |