BE AI: runs, proposals and scope
BE AI is the investigator. It reads history and asks machines read-only questions, then writes a conclusion. Anything that would change a machine is a proposal that a named person approves. It never executes anything itself.
What starts a run
- An alert-driven trigger: a check failing, a disk crossing 90%, a reboot. Each uses a recipe suited to it. The same machine and trigger run at most once every 30 minutes (
CLAUDE_DEDUPE_MIN). - The nightly pass: once a day per company at the time and days set on the company page, for every included online machine, one machine at a time. A machine that was healthy at its last pass with no events or alerts since is skipped and re-checked every three days. "Run pass now" runs it immediately.
- A question from Discuss, a follow-up on an executed action, or "verify the effect".
Each company has a daily cap (CLAUDE_DAILY_RUNS, default 50, overridable per company). Alerts, checks and approvals keep working when the cap is reached; investigations resume the next day.
What it may do
Read-only tools on the machine (diagnostics, metrics, processes, services, logs, checks, network facts, cellular on routers) and the hub's own history (heartbeats, events, check history). Budgets per run: 15 tool calls, 12 turns, three minutes; it must conclude. Everything a machine reports is treated as untrusted text; the model reports text that tried to instruct it.
Proposals
A proposal is a mutating tool with exact arguments and a reason: restart a service, end a process, run a bounded command, set checks. Approve, reject or retry from the approvals queue, the proposals page or inside the conversation; the result comes back in plain language and the output is kept with the proposal. Approving records your name. A machine refuses a shell command unless its own configuration allows it.
Conversations
Discuss opens a thread about one machine. Every message is a run whose context is the whole thread: your messages, the answers, and the output of every action executed in it. Ask for a change and it files specific proposals, approvable in the thread; their output feeds the next turn. A turn is started and then polled, so a hub restart mid-turn does not lose it.
Including and excluding
Per machine, on its notes tab: included in BE AI (off means never looked at, not even when asked), investigate alerts automatically, scheduled daily pass. Per company, on its page: BE AI on or off, auto-run on or off, runs per day, and the schedule. Notes on a machine or a company are given to the investigator as trusted technician instructions: what the machine is, maintenance windows, what never to propose.
Costs
The console estimates cost from token counts at the price configured on the hub (CLAUDE_PRICE_IN_PER_M, CLAUDE_PRICE_OUT_PER_M, CLAUDE_PRICE_CURRENCY). The Usage page shows machines, runs, tokens and estimated cost per company per month. Every run keeps its transcript, capped, so a conclusion can be traced to the tool calls that produced it.
Try it on one machine first.
The install line takes a minute. Uninstall the agent and you are out.