AgentCost – local CLI,attributes token cost in Claude Code/Cursor/Codex sessions

**AgentCost** ships as a local CLI on PyPI under the package name **agentcost-cli** (path version **0.1.0**). It attributes token cost across **Claude Code**, **Cursor**, and **Codex** sessions. The release was noted on Hacker News (item id **49034205**) with **2** points and **0** comments at the time of the summary, with the project page at pypi.org/project/agentcost-cli/0.1.0/.

As a **local CLI**, **AgentCost** runs on the machine where those agent sessions run rather than as a remote dashboard. Its stated job is cost attribution: map token spend back to individual **Claude Code**, **Cursor**, or **Codex** sessions so usage is not a single undifferentiated bill. The distribution channel is PyPI (**agentcost-cli**), which fits install into existing Python-based tooling paths.

For engineers who leave **Claude Code**, **Cursor**, or **Codex** running for long agent sessions, unattributed token cost is hard to audit. Session-level attribution turns a lump sum into a per-session signal: which runs burned budget, which tools dominate spend, and whether a workflow is cheap enough to keep automated. That matters for personal spend caps, team cost reviews, and deciding when to tighten prompts or tool calls.

The tool sits in a thin but growing layer around coding agents: **Claude Code**, **Cursor**, and **Codex** already own the session UX; cost visibility is often secondary or provider-console only. A local, multi-client attributer competes less with the agents themselves and more with ad hoc log scraping, spreadsheet export from provider dashboards, and one-off scripts. Early HN traction is thin (**2** points, **0** comments), so it is still a niche utility rather than a default in the stack.

Practical next step: install **agentcost-cli** from PyPI if you already run those three clients and need session-level token cost, then validate that attributed totals match what the provider shows for the same window. Watch whether attribution stays limited to **Claude Code** / **Cursor** / **Codex** or expands, and whether the local-CLI model is enough as teams move cost control into shared dashboards.