Tokenrail

Small software teams can run coding agents and model-backed products through their own provider keys without a consistent view of cost by task, pull request, ticket or customer. The supplied research confirms several live model gateways and a cost-aware rate-limit mechanism, while reporting no reviewed direct equivalent for a finance-facing monthly workload-cost statement. That output gap is bounded; established gateways already own the interception point.

Tokenrail would preserve organization, workspace, provider account, key reference, key-custody owner, gateway route, request time, model identity assertion, model version assertion, input and output usage, cached usage, tool usage, provider price-table version, estimated request cost, provider receipt, invoice line, task, agent identity assertion, pull request or ticket, customer-allocation assertion, attribution basis, shared-cost rule, budget policy, cap period, cap scope, exception authority, request decision, blocked-call event, fallback decision, destination acknowledgment, completed task assertion, monthly statement, finance review, approved allocation, correction and deletion as distinct records.

Usage metadata may be incomplete, provider prices can change and cached or batch calls may be billed differently. A tag linking a call to a ticket or customer is an allocation assertion, not proof that the work caused the cost or that the amount is billable. A blocked request is not savings unless an approved counterfactual supports it. Calling the report payroll implies an employment relationship and compensation that the product does not measure. Keys and prompts may contain secrets, customer data and source code; the gateway must minimize content, isolate tenants and never print key material.

The pilot should use synthetic prompts, valueless provider fixtures and non-production keys. The likely buyer is a technical founder, engineering-platform owner, finance lead or AI FinOps owner, but team size, provider mix, tagging discipline, latency tolerance, budget ownership and willingness to replace or layer over an established gateway remain unverified.

A technical founder, engineering-platform owner, finance lead or AI FinOps owner responsible for model-call budgets and reviewable engineering cost allocation.