Guide 58 of 61 · Running a fleet with Central Lite
Attribute AI spend by machine and person
Read your enrolled machines' model spend by day, machine, provider and model, and reconcile it against your providers' own usage and cost reports.
Built, not yet claimed preview: this guide describes a package that has not passed its release gate.
Read what your enrolled machines spent on model calls, per day, machine, provider and model, in the terminal and in the console; decide whether Central Lite should also keep spend per principal, and why it does not by default; then connect your vendor accounts so Central Lite reconciles the fleet's spend against the providers' own usage and cost reports.
1. Read the fleet's spend
Every enrolled machine uploads the model calls its proxy recorded, and Central Lite adds each one, once, to a row per UTC day, machine, provider, model and currency: calls, input and output tokens and the cost (a call left unpriced counts with none). Currencies are never added together. garm-central costs lists the rows of the last 30 days, or as many as --days asks; the host_id column is the machine's id, which garm-central node list pairs with its host name.
Only calls that go through a machine's proxy are counted, so a fleet whose agents have made no model call yet prints the header alone. Under the non-monitoring compliance profile a machine sends no model call to Central Lite at all, only hourly counts, so its spend never reaches these rows.
garm-central costs --days 7
garm-central node listChecked by the walk: this command ran as written on a Central Lite host.
See also
See what each task costRoute an agent's model calls through AgentGarm so each one is counted and priced, and read the spend by provider and model.
2. Open Costs in the console
In the console, Costs shows the same spend with a Range (Last 7, 30, 90 or 365 days) and a Group: By model, By vendor account (reconciled), By principal or By credential; choose Show costs. The Summary gives the cost, calls and tokens in one currency (any other currency is listed beside it, never added), the Daily trend draws a pair of bars per UTC day, or per week for a range over 31 days (ingested and reconciled cost; Reconciled and Priced locally under By vendor account), and Rows lists each day with its Reconciliation status: within tolerance, drift, no workspace cost, failed or not reconciled. Export CSV saves the rows loaded, with amounts as plain numbers and the currency in a column of its own.
What you can group by depends on your role and its scope. By principal and By credential need a role that may read people's spend (owner, admin, security, auditor or finops, not viewer) bound at org scope, and By vendor account needs a role at org scope. A role scoped to a host group or to hosts sees By model only, summed over the machines it covers (Give people the right access in the console).
By model sums the rows of every machine you can see, so the console shows no per-machine split; the terminal does.
3. Decide whether to keep spend per person
Each call names a principal: the one whose listed credential it presented (a gateway key, or another key registered with identity credential add), a delegated principal from an access token, or else the principal its task ran as, the one Central Lite's registry binds to the user's login. Machines upload those ids only as keyed hashes, and Central Lite turns back into registry ids the ones it authored. Central Lite keeps no row per principal and credential until you run garm-central settings principal-costs on, a chained admin action; until then costs --by-principal prints a note that the rows are off, and the console says the same under By principal and By credential.
It is off by default because the ingest volume of these rows is measured against a budget that is still provisional: it stays off until a replay of real recorded machine-days through Central Lite passes it. Central Lite adds these rows only for calls it receives while the setting is on; nothing earlier is filled in, and off stops adding them.
garm-central settings principal-costs on
garm-central costs --by-principal --days 7garm-central settings principal-costs offChecked by the walk: these commands ran as written on a Central Lite host.
See also
Prepare a works-council or DPIA packGenerate templates that describe what your deployment collects, keeps and sends, and switch the fleet to the non-monitoring profile during talks.
4. Connect your vendor accounts
Central Lite reconciles the fleet's calls against a provider's own usage and cost reports with a least-privilege admin (or analytics) key for the account: anthropic, openai, cursor, devin or github-copilot, one account per provider. For Anthropic the account is the organisation's UUID, for OpenAI the organisation id (org-…). garm-central connector add reads the key from stdin only and seals it in Central Lite's keystore; it is never stored in the database, printed or logged, and connector rotate replaces it the same way.
Once a machine enrols, it refuses to reconcile on its own, so a fleet reconciles in Central Lite only, and connector add refuses an account an enrolled machine still has a key for. serve reconciles every hour (--reconcile-every) each of an account's last three settled days that has no run yet or whose run failed; connector run reconciles one day now, by default the last settled one. A failed run changes no spend, raises a reconciliation_failed alert and exits 1. Central Lite refuses to run one as root.
read -rs KEY && printf '%s' "$KEY" | garm-central connector add anthropic --account <organisation-uuid>; unset KEYgarm-central connector run anthropicgarm-central connector listChecked by the walk: this command ran as written on a Central Lite host.
See also
Reconcile spend with provider invoicesReconcile a day of this machine's model calls against your provider's own usage and cost reports, read where they differ, and export FOCUS rows.
5. Compare your spend with the provider's
garm-central costs --reconciled lists per day, provider and model the fleet's calls and cost beside the reconciled account, its status (WITHIN_TOLERANCE, DRIFT, NO_WORKSPACE_COST, FAILED, or not reconciled) and the local, report and adopted figures, with - where a side has none. A model only the provider reported still gets its row.
--by-key goes per vendor API key and model: the requests and their usage priced locally, which is not the billed cost, joined to the credential you listed with that key's --issuer-key-id and so to its principal. A minute two credentials both held reads shared key, ambiguous. A key the provider billed after you revoked its credential raises revoked_credential_billed in garm-central alerts list.
Central Lite itself enforces no spend limit: a budget lives in the policy you publish, and each machine holds every task to it.
garm-central costs --reconciled --days 7
garm-central costs --reconciled --by-key --days 7Checked by the walk: this command ran as written on a Central Lite host.
See also
Set spend budgets and stop runaway loopsGive each agent task a cost budget and a token rate, see what the agent meets at the cap, and stop loops and runaway subagents with limits of their own.