AI Gateway

Best AI Gateway to Monitor Claude Code Token Usage

Best AI Gateway to Monitor Claude Code Token Usage
The best AI gateway to monitor Claude Code token usage is Bifrost. Get per-request cost attribution, per-team budgets, Prometheus metrics, and OTel tracing.

Claude Code spend now averages $150 to $250 per developer per month at scale in enterprise rollouts, according to CloudZero's analysis of Anthropic deployment data, and a single subagent-heavy session can reach four-figure costs in a few hours. Most engineering teams have no centralized visibility into which models, developers, or sessions drive that consumption. Bifrost, the open-source AI gateway by Maxim AI, is the best AI gateway to monitor Claude Code token usage. Built in Go and available on GitHub under Apache 2.0, Bifrost sits between Claude Code and any LLM provider and captures per-request token counts, per-virtual-key cost attribution, and full session traces with 11 microseconds of overhead at 5,000 requests per second.

Why Claude Code Token Usage Demands Centralized Monitoring

Claude Code token consumption is unpredictable at team scale because the same product can be invoked in radically different ways. A focused bug fix may use 20,000 tokens. A multi-file refactor with subagents can use millions. Anthropic's Economic Index reports that a 1% increase in input length is associated with a 0.38% increase in output length, meaning token costs compound non-linearly as context grows.

Without monitoring at the gateway layer, engineering leaders cannot answer basic questions:

  • Which teams or developers account for the majority of monthly Claude Code spend?
  • How much of that spend is on Opus versus Sonnet versus Haiku?
  • Which sessions are burning tokens through autocompact loops or runaway subagent fan-out?
  • Are projected budgets being breached, and by which cost center?

Claude Code's built-in /cost command and the Anthropic Console show consumption at the session and account level, but they cannot attribute spend across multiple developers, enforce per-team budgets, or feed data into Prometheus, Grafana, or Datadog. That is the job of an AI gateway placed in front of Claude Code, where every request is logged, costed, and attributed before it ever reaches Anthropic. The Bifrost AI gateway is built for this role.

What an AI Gateway Should Track for Claude Code Token Usage

An AI gateway used to monitor Claude Code token usage should capture and expose the following data points at request granularity:

  • Per-request token counts: input tokens, output tokens, and cache reads or writes for every Claude Code call
  • Per-request cost: dollar amount computed against current Anthropic, Bedrock, Vertex, or Azure pricing
  • Per-user and per-team attribution: which developer, team, or business unit issued the request
  • Model breakdown: how much spend goes to Opus, Sonnet, and Haiku separately
  • Session and trace correlation: full session view across multi-turn Claude Code interactions
  • Budget enforcement: ability to block requests when a team or developer exceeds an allocation
  • Standards-based export: Prometheus, OpenTelemetry, and Datadog integration so the data lives where the rest of engineering monitoring lives

A gateway that captures only aggregate totals does not solve the team-scale problem. The gateway has to attribute spend down to the individual key or developer, and it has to make that data extractable into existing observability platforms. The Bifrost governance model is built around exactly this pattern: virtual keys as the unit of attribution, hierarchical budgets as the unit of enforcement, and standards-based export as the unit of integration.

Bifrost: The Best AI Gateway to Monitor Claude Code Token Usage

Bifrost is built around four monitoring capabilities that map directly to the requirements above.

Per-request cost and token attribution. Bifrost logs every Claude Code request with input tokens, output tokens, latency, model, provider, status, and computed cost. The pricing catalog is automatically synced from a remote datasheet, and custom pricing overrides let teams reflect negotiated Anthropic, AWS Bedrock, or Vertex rates. Cost is computed at request time, not in a nightly batch.

Per-team and per-developer visibility through virtual keys. Virtual keys are the primary governance entity in Bifrost. Each developer or team gets a virtual key that Claude Code uses as its ANTHROPIC_AUTH_TOKEN, and every request is automatically tagged with that key. Dashboards, metrics, and exports filter by virtual key, team, or customer, producing a clear answer to who spent what.

Native Prometheus metrics and OpenTelemetry tracing. Bifrost exposes a /metrics endpoint that Prometheus can scrape or push to a Push Gateway for multi-node clusters. The same data also flows through OpenTelemetry as span-level traces, exportable to Grafana, New Relic, Honeycomb, or any OTLP-compatible backend. Metrics collection runs asynchronously with no impact on Claude Code request latency.

Real-time built-in dashboard. Bifrost ships with a built-in dashboard at http://localhost:8080/logs that shows token usage, cost, model breakdown, and latency, filterable by virtual key and time window. No external tools are required to see who consumed what. The same dashboard supports per-team filtering for engineering leads who need a weekly review without setting up Grafana.

For teams that already run Datadog, Grafana, New Relic, or Honeycomb, the OTel observability plugin pushes APM traces and LLM Observability data directly into each platform.

How Bifrost Connects to Claude Code

The Claude Code integration requires changing two environment variables in Claude Code's settings.json and creating a virtual key in Bifrost.

"env": {
  "ANTHROPIC_BASE_URL": "http://localhost:8080/anthropic",
  "ANTHROPIC_AUTH_TOKEN": "your-bifrost-virtual-key"
}

After this change, every Claude Code request routes through Bifrost transparently. Developers continue using Claude Code exactly as before, with no workflow changes. The open-source Bifrost gateway translates between the Anthropic API format and any target provider, including AWS Bedrock, Google Vertex AI, Azure, OpenAI, Groq, and Mistral, so platform teams can route Claude Code to whichever provider holds the contract or fits the cost model. The ANTHROPIC_AUTH_TOKEN method does not require an Anthropic account login at all; the virtual key handles authentication and billing routing on its own.

Once routing is in place, monitoring is automatic. No additional SDK, instrumentation, or code change in Claude Code is needed.

Budget Enforcement and Active Cost Control

Visibility without enforcement leaves cost overruns to discover after the fact. Bifrost moves from passive monitoring to active control through hierarchical budgets attached to virtual keys, teams, and customers. Each level enforces independently.

  • Per-developer daily limits prevent a single runaway session from burning the team budget
  • Per-team monthly budgets keep cost predictable at the engineering manager level
  • Organization-wide caps give finance and platform leads a hard ceiling

When a budget is exhausted, Bifrost blocks further requests automatically. Calendar-aligned reset durations (1d, 1w, 1M, 1Y) align with billing cycles. Rate limits on tokens per hour and requests per minute add another layer of protection against accidental burst spend, especially for subagent-heavy Claude Code workflows that have been known to fan out to dozens of parallel sessions.

For platform teams, the governance resource covers the full virtual-key, RBAC, and SSO model. For deeper monitoring patterns, the MCP gateway resource walks through how Code Mode and tool filtering further reduce Claude Code token consumption.

How Bifrost Compares to Alternative Monitoring Approaches

Engineering teams evaluating how to monitor Claude Code token usage typically consider four options.

Claude Code's built-in /cost command: Useful for individual developers checking their own session. It cannot aggregate across users, attribute to teams, enforce budgets, or feed Prometheus.

Anthropic Console organizational view: Adds workspace-level cost tracking and per-user breakdowns when developers are on API billing. It does not enforce budgets at request time, does not stream metrics to Prometheus or OTel, and only reflects spend from Anthropic-hosted models, not Bedrock, Vertex, or Azure deployments of Claude.

LiteLLM proxy: A Python proxy that supports routing to multiple providers including Anthropic. It logs token usage and cost per virtual key with a PostgreSQL backend. The architecture introduces a Python runtime baseline latency that is higher than Go-based gateways and requires an external database for spend tracking. Teams comparing options can review Bifrost as a drop-in LiteLLM alternative for the complete feature delta.

The Bifrost AI gateway: Provides the most comprehensive monitoring stack for Claude Code token usage: per-request cost, virtual key attribution, hierarchical budgets, Prometheus and OTel export, a built-in dashboard, and a Datadog connector. Bifrost adds only 11 microseconds of overhead at 5,000 requests per second in published benchmarks, so the monitoring layer does not slow down interactive Claude Code sessions. The Go-based architecture handles governance enforcement, cost logging, and observability export concurrently without impacting request latency.

Getting Started with Bifrost for Claude Code Token Monitoring

Engineering organizations running Claude Code at team scale need an AI gateway that monitors token usage at request granularity, attributes spend to virtual keys, enforces hierarchical budgets, and exports data to existing observability platforms. Bifrost delivers all four in a single open-source binary that deploys with zero configuration.

For regulated industries and enterprise teams that need air-gapped deployment, VPC isolation, SSO integration with Okta or Entra, or RBAC for finance and platform teams, Bifrost Enterprise adds clustering, vault support, audit logs, and adaptive load balancing on top of the open-source core.

To see how Bifrost can centralize Claude Code token usage monitoring for your engineering organization, book a demo with the Bifrost team.

Read next