Bring LiteLLM gateway spend into Cloud Intelligence™

Run one small exporter next to your LiteLLM proxy and every model call shows up as labeled spend, with history backfilled.

If you route LLM traffic through a self-hosted LiteLLM gateway, you already have some of the richest cost data in your company: every call metered, priced, and tagged with the key, team, and end customer that made it. The problem is where that data lives. It sits in the proxy's own Postgres, invisible to the place where you analyze everything else you spend. So AI cost reviews turn into SQL against the gateway database while the rest of the bill lives in your FinOps tooling.

We wanted to close this gap. Cloud Intelligence™ now ingests LiteLLM spend, one event per model call, into the same reports as your cloud bill.

What you get

You run the litellm-datahub-exporter next to your proxy: one static binary or container, configured with your proxy URL and two keys. It polls LiteLLM's spend APIs and pushes labeled events to the DataHub API. Setup is deliberately boring.

Each call arrives labeled with provider, model, virtual key, team, and, if you pass them through LiteLLM, your feature name and end-customer identifier. Those become unified dimensions in reports, allocations, budgets so cost per feature per customer is a report you build in a minute, and the same events light up the AI Intelligence dashboard next to your Anthropic, OpenAI, Bedrock, Gemini, Cursor spend full list of integrations.

Three design choices that matter

There is no pipeline. No OpenTelemetry (OTEL) collector to deploy, no filter rules to maintain, no telemetry backend between your gateway and your reports, and nothing for DoiT to reach into your network for. The exporter is the whole integration, and it sits outside the request path, so it can never add latency to a model call or drop one.

History comes with you. On first run the exporter backfills from your spend-log retention, up to two years. Approaches that tap the gateway's live telemetry stream only see traffic from the day you instrumented; this one starts with the past.

It survives real operations. Every event has a deterministic id, and re-exports overwrite instead of duplicating. Restart the exporter, lose its checkpoint, replay a whole month after an outage; the numbers stay right.

LiteLLM computes cost from its price map, not from provider invoices, so every event is labeled as estimated spend. Because Cloud Intelligence™ also has native Anthropic and OpenAI integrations that pull provider-side cost and usage reporting, you can reconcile the gateway's view against what the providers actually charge, in one report.

What never leaves your network

The exporter is open source under Apache-2.0, small enough for your security team to read in a sitting, and it runs entirely in your infrastructure. Your LiteLLM admin key is used only against your own proxy. Spend rows pass through a strict field allowlist, so prompt and response content cannot be exported even by misconfiguration; that guarantee is enforced in code you can audit, not in a filter you have to remember to configure. Releases ship as signed containers and binaries with published SBOMs, built by a public CI workflow you can verify with cosign.

Get started

1. Create a Cloud Intelligence™ API token with the DataHubAdmin scope.
2. Run the exporter beside your proxy with Docker, Helm, or systemd.
3. Label your traffic with feature and end-customer identifiers, then open Cloud Analytics and group by them.

Your spend appears under the LiteLLM provider about 15 minutes after the first cycle. The integration needs a DataHub subscription on your Cloud Intelligence™ tier; the LiteLLM connection guide covers the rest.