Handing a language model the power button

Lightning talk · 95 seconds

Handing a language model the power button

mcidrac lets you talk to a rack of Dell PowerEdge servers. The interesting part isn’t the talking — it’s what had to become true before that was a good idea.

▶Listen to the talk95sGet started →

1. 01the tax

   Four tabs to read one temperature

   Something is wrong with a server, and the server is somewhere else — another room, another building, another state. So you open the iDRAC web console, wait for it to load, and click through four tabs to find a single inlet temperature. One box under a desk or four hundred in a colo, the tax is the same, and we have all been paying it.

2. 02the sentence

   Replace all of that with a sentence

   “Boot r740-lab from the rescue ISO.” The ISO mounts, a one-time boot flag is set to the virtual CD, the host restarts. Twenty-seven tools cover power, sensors, the hardware event log, inventory down to RAID volumes, BIOS attributes, and full configuration backup — each one a sentence instead of a session.

3. 03the catch

   What does “it worked” actually mean?

   Handing a language model the power button raises a harder question than convenience. A baseboard management controller will happily return 200 OK for a change it silently rejected. Dell validates BIOS settings minutes later, during boot. A virtual media mount can be accepted and then never happen. If a model chains its next destructive action on the word “applied”, that gap is where the bad day lives.

4. 04the rule

   Read it back, or it didn’t happen

   So there is one rule, and everything else follows from it. A step is not okay because the request was accepted. It is okay because the state read back correct. Mount the ISO, then re-read the device. Set the boot flag, then re-read the flag — and refuse to reboot the machine until it verifies. When a sequence fails halfway, the error carries a ledger of what already changed, because that is the information you need to recover.

5. 05the lab

   Six fake iDRACs, two generations, no hardware

   To prove any of that, the project ships a laboratory. Dell publishes real Redfish captures of their own machines; the harness serves six of them as working fake BMCs, plus a behavioral simulator for the write paths — power states, job queues, read-only attribute rejections. The whole tool surface runs green before it touches your rack.

6. 06the payoff

   It caught a real one

   iDRAC 10 deleted a virtual media path that every previous generation had. Ninety tool calls across the fake fleet found it in about a minute — on a fake server, for free, instead of on your real one. That same lab is now the honest way to answer a newer question: is my local model good enough at tool calling to trust with hardware?

And if the whole point is keeping secrets in the building, the model can run there too — mcidrac with a local model on a DGX Spark keeps the credentials, the telemetry, and the reasoning inside one boundary you control.

Try the fake fleet ·All 27 tools ·The read-back rule