Bifrost gateway AI attribution in Cloud Intelligence™

See per-request LLM spend by virtual key, team, and customer from your self-hosted Maxim AI Bifrost gateway, next to your cloud bill.

If you route LLM traffic through a self-hosted Bifrost gateway, you already have the attribution problem half-solved. Every request passes through a virtual key that belongs to a team or a customer, and the gateway prices each call from its model catalog. The trouble is where that data lives: inside your VPC, in the gateway's own log store. Your provider bills tell you how much you spent on Anthropic or Bedrock last month. They can't tell you which team, which customer, or which application drove it. The gateway knows, and until now that knowledge stayed home.

The bifrost-datahub-exporter ships your gateway's request data into Cloud Intelligence™.

What you get

One small container or binary runs next to your gateway. It polls Bifrost's logs API incrementally and pushes one event per request to Cloud Intelligence™, labeled with provider, model, virtual key, team, and customer, straight from the gateway's governance hierarchy. Those labels become first-class dimensions in Cloud Analytics: group a report by team, build an allocation per customer, set a budget on a virtual key, point anomaly detection at any of it. The spend also lands in the AI Intelligence dashboard alongside your other AI providers, with prompt and completion token counts and, on gateway v2.0.0 and later, the input/output cost split.

One thing to keep straight: Bifrost bills nothing. You bring your own provider keys, so the invoiced spend already reaches Cloud Intelligence™ through your provider and cloud cost feeds. The gateway's numbers are catalog-priced estimates, and the integration treats them that way. Every event carries a cost_basis: estimated label. Use this dataset to answer who and what; your billing data still answers how much.

The exporter is open source (Apache-2.0), built for the security review it will get before touching your VPC. It decodes gateway logs through a strict field allowlist, so prompt and response content has no code path out of your network, and a leak-canary test in CI keeps it that way. Gateway credentials stay inside; the only outbound call is to the Cloud Intelligence™ APIs. Releases ship as reproducible builds with cosign signatures, so you can verify what you run instead of trusting us.

Setup is deliberately boring: two environment variables, a Helm chart or docker-compose entry, done.

Get started

1. Create a Cloud Intelligence™ API token with the DataHub Admin permission.
2. Run the exporter next to your gateway with your gateway URL and the token. The Bifrost connection guide has the Helm, docker-compose, and systemd options.
3. Within about 15 minutes, your data appears under DataHub > Datasets > Bifrost and in Cloud Intelligence™ as the Bifrost provider.

If you run Bifrost in production, we want to hear what your reports look like at real traffic volumes: which labels carry your allocation model, and what breaks first. That feedback shapes what we build next here.