How AxiGate works

Your agents send their AI calls to AxiGate instead of straight to OpenAI, Anthropic or Gemini. AxiGate passes each prompt on unchanged, with your agent's own provider key, and hands the answer straight back.

On the way, it notes what the call cost. Once a limit is reached, it refuses the next call before it costs anything.

The gateway runs on your own computer or server, so your agents' calls go from there straight to the provider.

A run is one job an agent does, from start to finish. Its calls are counted together, so a limit can stop one stuck job while the others carry on.

Start an agent with axigate-finops run and each start is a run of its own. When the agent finishes you get a receipt: how many calls it made, what they cost, and whether a limit stopped it.

Through a gateway you leave running (axigate-finops serve or axigate-finops gateway), each Claude Code session counts as a run with nothing to set up. The Runs page lists the costliest runs and any a limit stopped, with what each cost; open one to see each call it made.

For each call AxiGate keeps a record: the times, the provider, the model and program names, the tokens, the cost, the outcome, your labels, run and session ids, and a scrambled key id ending in the key's last 4 characters. A refused call is kept too, so a stop never leaves a gap in the record.

Prompts go to the provider and answers come back unchanged, and neither is recorded. Your provider key is passed through and never stored; AxiGate tells keys apart by a fingerprint, not by the key itself.

The record is a file on the computer running the gateway, which prints where it is when it starts. With a hosted workspace, the gateway also sends that same record there, and nothing more.

Either way, the gateway runs on your machines. What differs is where you read the results.

AXIGATE_INGEST_KEY=<the key from Connect a gateway> axigate-finops gateway --ingest-url https://<your workspace>/api/ingest

How far back you can look: all of it on the free edition, 14 days on Hosted Free, 90 days on Team and 13 months on Business. Nothing older is deleted; it is just outside the view.

Which plan does what compares them side by side, and Keep a year of history explains how far back each plan looks.

Every figure says how sure it is. A call's cost starts as an estimate: AxiGate prices the tokens it used from a price list for that model, together with what the provider's answer says about the price: the speed tier that served the call, its cache writes, its web searches, and whether its prompt was long enough for the long-context rate. See where the money goes lists them.

The provider's own bill is the better number. Bring in an OpenAI or Anthropic bill and AxiGate puts it next to what was recorded, with any difference on its own line.

A bill says what the month cost, not which agent spent it, so the split by team and agent stays an estimate. A difference is never spread across teams to make the totals add up.

Close the month against the invoice shows how to bring in the bill and confirm a month.