A fresh entry hit PyPI this week that should catch the attention of anyone running AI-assisted development at scale. The agentcost-cli package, released as version 0.1.0, provides a local command-line interface for attributing and tracking token consumption across Claude Code, Cursor, and Codex sessions. Unlike cloud-based billing dashboards from Anthropic or OpenAI, this tool runs entirely on your own machine—no API calls to track usage, no vendor lock-in on cost analytics.

What AgentCost Actually Does

The CLI intercepts session metadata from supported AI coding tools and logs token usage locally. For teams running multiple developers on Cursor Enterprise seats or Claude Code instances, visibility into per-session consumption has traditionally required either manual calculation using API pricing or reliance on each vendor's proprietary tracking. AgentCost sidesteps that by capturing the data at the client level, outputting structured logs you can pipe into your own monitoring stack. The README notes support for all three major AI coding assistants, though early adopters have reported varying levels of session metadata extraction depending on tool version and configuration.

Why This Matters for Developer Economics

Token costs aren't trivial anymore. A single sprint involving heavy Claude Code usage can generate thousands of dollars in API calls, especially when developers leave sessions running or use verbose prompting strategies. Until now, the granularity to understand which features, repos, or team members were driving the highest consumption required either enterprise tier subscriptions from the AI providers or custom instrumentation. AgentCost represents a shift toward developer-controlled observability—treating AI usage like any other infrastructure metric that engineering teams should own and optimize.

The Current State of Affairs

The package is fresh, sitting at 0.1.0 on PyPI, which means expect rough edges. Hacker News comments noted the tool's utility but flagged that session attribution accuracy depends heavily on how each supported assistant exposes internal telemetry. Cursor and Codex have different logging formats than Claude Code's native output, so users may need to tweak configuration for their specific setup. That said, the core value proposition—local-first token cost tracking without vendor dependency—is exactly the kind of infrastructure tooling the AI-assisted development ecosystem needs as costs scale upward.

Key Takeaways

  • AgentCost CLI runs locally with no cloud dependencies or API calls back to vendors
  • Tracks token usage across Claude Code, Cursor, and Codex sessions simultaneously
  • Outputs structured logs compatible with existing monitoring pipelines
  • Version 0.1.0 indicates early-stage software—expect active development and potential breaking changes
  • Particularly valuable for teams managing multiple developer seats or enterprise agreements where granular cost attribution matters

The Bottom Line

We're entering the phase where AI coding assistants aren't just productivity tools—they're line items on engineering budgets, and tools like AgentCost are the first wave of infrastructure to manage that reality. If you're running more than a couple of these tools in your org, you should already be thinking about observability. This is a solid starting point.