Architecture
A lightweight, transparent proxy sits between your application and every AI provider. Total financial visibility.
Architecture overview
TOLVYN intercepts every AI request routed through it, before it reaches the provider — recording, attributing, and enforcing limits in real time.
Step by step
Point your AI SDK's base URL at TOLVYN's proxy. Our own SDKs read it from TOLVYN_PROXY_URL; with a provider's official SDK you set that SDK's own base-URL option. No SDK changes, no code refactors, no downtime. Works with any HTTP-based model API.
Register your provider key with TOLVYN once. It is stored encrypted, never in plaintext, and your application then sends a TOLVYN key instead — TOLVYN substitutes the provider key on the way out. Your provider key stops living in your application's configuration, and rotating it is one change in one place. TLS everywhere, end to end.
Every request is received by TOLVYN's proxy layer. The request metadata — model, token counts, latency, status — is captured and appended to the tamper-evident ledger.
Each ledger entry is hash-chained to the previous, making it cryptographically tamper-evident. Any retroactive modification of a record breaks the chain — immediately detectable.
Tag requests with arbitrary metadata — team, service, environment, feature — via HTTP headers or the TOLVYN SDK helper. TOLVYN uses these tags to break down costs at any granularity you care about.
See which team spent $2,400 on GPT-4o last week, which microservice is driving token growth, and which model is delivering the best cost-per-output ratio — all in one dashboard.
Set hard spending limits per team, per service, or per model. Define alert thresholds at 50% (info), 75% (warning), and 90%/100% (critical) of budget. TOLVYN can automatically block requests once a budget is exhausted — no surprise invoices at month end.
Finance teams get an audit trail of everything the proxy metered, plus reconciliation against the provider invoice to surface anything that did not go through it. Engineering teams get guardrails. Everyone stays aligned on AI spend without slowing down development velocity.