See every request. Priced, owned, traced.
Point your agents at one gateway. Every call is recorded, priced at public catalog list, and kept as a trace you can open.
Your AI spend is an invoice. No line items, no owner.
Correct, and it explains nothing. No line for the retry loop that ran all weekend.
Your provider does not know which agent made the call, or which workflow it belonged to.
Without an owner and a trace, spend is a debate rather than something anyone can act on.
Change one base URL. Everything after is on the record.
- one base url
- The gateway is a transparent reverse proxy for supported providers. You change the base URL. Your agents do not change.
- every request recorded
- Provider-reported usage per call: model, input and output tokens, cache reads and writes, status and latency.
- priced at public list
- Each request carries a public-catalog list-price subtotal. It is a subtotal, not your invoice.
- an owner for every number
- Activity groups requests by the people who ran them and the work they were doing.
- byte-safe
- Cache hints are added to the upstream request only. What the provider returns is what your agent gets.
One request, all the way down.
A trace is the request itself, not a summary written after it. Nothing is reconstructed.
- spans
- each step, with its own duration
- tokens
- in, out and cached, as reported
- owner
- the agent and the workflow
- payload
- what went out, and what came back
- identity
- provider, model, status, request id
Every number has an owner. Or it is marked unattributed.
Spend is grouped by the work it did and the people who ran it, so a line on the chart resolves to a workflow, a key, an agent — or to nothing, which is also an answer.
What cannot be attributed is shown as unattributed rather than folded into the nearest owner.
Priced at public list. Never at a number we made up.
- catalog list price
- Spend is provider-reported usage costed against the public model catalog. It is a list-price subtotal, not your bill, and it is labeled as one everywhere it appears.
- unpriced stays unpriced
- A model the catalog does not price is shown as unpriced. It is never quietly counted as $0.
- $0.0000 until proven
- Verified savings start at zero and move only on provider-causal evidence. Local estimates are labeled inferred and never promoted.
Spend shown across this page is a public-catalog list-price subtotal from provider-reported usage, not an invoice.
What powers it.
Stop guessing
what it cost.
Caveman Platform is in private development. Screens are real product UI from a seeded demo workspace. Spend shown is a public-catalog list-price subtotal from provider-reported usage; unknown models stay unpriced, and verified savings start at $0.0000. The local Proxy is free and needs no account.