Anthropic previews /usage token breakdown for Claude Code Skills, Agents, MCPs, and Plugins

• Anthropic previewed a Claude Code /usage command that breaks token spend down by Skills, Agents, MCPs, and Plugins, according to bcherny's preview.
• The feature is shipping in the CLI first, while the same preview says Desktop support is planned next.
• The breakdown targets a real blind spot in agentic Claude Code setups, because token burn gets harder to attribute once Skills, MCP servers, and delegated agents all join one run, as bcherny framed it and aakashgupta's workflow thread illustrates.
• There is already a stopgap for skill-specific inspection, because Everlier's reply points to skilled, an open source tool for digging further into skill usage.

Anthropic only showed one line of product surface here, but it is a useful one. You can see the initial preview, jump straight to the community-made skilled repo, and compare that with one power-user workflow built around dynamically loaded skills and parallel sub-agents.

The new command is simple: /usage will show which parts of a Claude Code run are actually consuming tokens.

That matters because Claude Code sessions now span more than one prompt window. A single task can pull in a Skill, call an MCP server, hand work to an Agent, and invoke a Plugin. Without attribution, the cost lands as one blended number.

The wording in bcherny's post is also specific about scope. This is not just a session total. It is a component-level breakdown across four extension surfaces that users can actually swap, add, and tune.

One reason this preview landed fast on X is that skills-heavy Claude setups are already getting elaborate. In aakashgupta's workflow thread, he describes a system that spreads work across Cowork, Claude Code, and Dispatch, then loads domain-specific skills only when a task matches.

That thread includes three details that make a usage breakdown more useful than a plain token counter:

• Skills are loaded dynamically, not all at once, per aakashgupta's workflow thread.
• The setup uses parallel sub-agents across different surfaces, also per the same thread.
• Gupta's separate post argues many paid Claude users never touch features like file access, parallel agents, or persistent workflows, according to aakashgupta's capability-gap post.

Once a workflow starts composing those pieces, the expensive part is not always the base chat itself. It can be the scaffolding around it.

Anthropic's preview also surfaced a small ecosystem tell. Within minutes, Everlier's reply pointed users to skilled, a GitHub project for inspecting skill usage more closely.

That suggests the visibility problem was already sharp enough for third-party tools to appear before Anthropic shipped its own first-party view. The official version goes broader by naming Skills, Agents, MCPs, and Plugins together, while the community tool focuses on the Skills slice.

The rollout note in bcherny's preview is easy to miss: the breakdown is coming to the CLI first, with Desktop later.

That split says something about where the pain showed up first. The users stacking Skills, Agents, MCPs, and Plugins into one trace are still mostly living in Claude Code, not a lighter desktop chat surface.