Where do your Claude Code tokens go?
Run /usage in Claude Code to see skills, subagents, plugins and MCP servers for this machine. Why usage climbs, and how to see it across every PC.
Run /usage in Claude Code. On Pro, Max, Team and Enterprise it attributes your recent usage to skills, subagents, plugins and individual MCP servers, and flags long context and cache misses when one of them is 10 % or more of recent usage; press d or w for the last 24 hours or 7 days (Commands). The figures come from the session history of the machine you run it on, so usage from your other computers and from claude.ai isn't included.
What does /usage show?
/usage shows the session cost, your plan's usage limits and activity stats; /cost and /stats are aliases of it. On Pro, Max, Team and Enterprise it adds the breakdown by skill, subagent, plugin and MCP server, behaviour flags such as long context and cache misses when they pass 10 % of recent usage, and rows for loops (Commands). The dollar figure in its Session block is meant for API users: on Pro and Max it isn't what you pay.
/usage: your plan's limits and where recent usage went
Example output. The bars are your account's limits; the breakdown under them only counts this machine's history.
| Command | Shows |
|---|---|
| /usage | your plan's limits and what used them |
The docs call the figures approximate. They are computed from the local session history, so a second laptop, a desktop at the office or your chats on claude.ai don't appear, even though they draw on the same limits (Commands).
Why does Claude Code usage climb so fast?
Anthropic's cost guide lists the usual causes (Manage costs effectively):
- the full conversation is sent with every request,
- cache misses after a break longer than the cache lifetime,
- scheduled tasks firing while you're idle,
- messages between sessions and goal check-ins,
- agent teammates,
- compaction of a large context.
The first one matters most in long sessions: each request carries everything before it, so the cost per turn keeps growing.
Each request re-sends the conversation, so late turns cost the most
Tokens sent to the model at each turn of one chat (illustrative: 200-token messages, 600-token replies, 3k of attachments).
- Attachments
- History sent again
- Your message
- Reply
Takeawaya new chat for each new topic brings the cost back to turn 1.
| Turn | Tokens sent |
|---|---|
| 1 | 3.8k |
| 2 | 4.6k |
| 4 | 6.2k |
| 6 | 7.8k |
| 8 | 9.4k |
| 10 | 11k |
| 12 | 12.6k |
| 14 | 14.2k |
| 16 | 15.8k |
| 18 | 17.4k |
| 20 | 19k |
Why do subagents and agent teams cost more?
Subagents and agent teammates do work of their own, and all of it counts toward the same limits. Anthropic's guide gives one order of magnitude: agent teams use approximately 7x more tokens than standard sessions when teammates run in plan mode (Manage costs effectively). The subagent rows of /usage show how much of your recent usage went there.
What is a cache miss, and why does a break cost tokens?
Claude Code caches the start of the conversation so the next request can reuse it at a lower cost. On a subscription, the main conversation's prompt cache lives one hour; once you draw on usage credits it lives five minutes, unless you set promptCacheTtl to 1h (Claude Code v2.1.242 or later) (Prompt caching). Come back after the cache has expired and the next request reads the whole context again at full price. That is the "cache misses" flag in /usage.
Claude Code and your weekly limit
Anthropic is explicit: "your usage of all different Claude product surfaces (claude.ai, Claude Code, Claude Desktop) counts towards the same usage limit" (How do usage and length limits work?). If a big part of your week goes to Claude Code, it is worth knowing which part of Claude Code.
The week split by product, in points of the weekly limit
Example: the week at 63% of the limit, day 4 of 7. The white marker is where a steady pace would be.
- Claude Code60%38 pts
- Chats27%17 pts
- Cowork10%6 pts
- Other3%2 pts
At this pacethe limit would be reached around day 6.3 of 7.
| Product | Share of the usage | Points of the limit |
|---|---|---|
| Claude Code | 60% | 38 |
| Chats | 27% | 17 |
| Cowork | 10% | 6 |
| Other | 3% | 2 |
How to see the breakdown across every PC
UsageMeter Link, our free Windows app, reads Claude Code's transcripts on each PC and adds them up on your UsageMeter dashboard. Version 0.5 adds the breakdown: the share of subagents, MCP servers, skills and plugins, the tokens re-read after a break and your limit hits, over 7, 30 or 90 days and across all your PCs, with one concrete tip per line (Pro). A Sessions tab stays on your PC. Project and server names are hashed unless you allow them.
It doesn't replace /usage: it answers the questions /usage can't, across machines and over weeks.
At a glance
/usage in Claude Code | UsageMeter Link 0.5 | |
|---|---|---|
| Skills, subagents, plugins, MCP servers | Yes (Pro, Max, Team, Enterprise) | Yes (Pro, on the dashboard) |
| Cache misses, long context | Flagged at 10 % or more | Tokens re-read after a break |
| Period | Last 24 hours or 7 days | 7, 30 or 90 days |
| Machines | This one | Every PC with the app |
| Usage from claude.ai | No | With the browser extension, on the dashboard |
Sources: Commands; UsageMeter's changelog.
FAQ
How do I see which MCP server uses the most tokens in Claude Code?
Run /usage on a Pro, Max, Team or Enterprise plan: it attributes recent usage to individual MCP servers, as well as skills, subagents and plugins, for the machine you're on. Press d or w to switch between the last 24 hours and 7 days.
Why doesn't /usage match what claude.ai shows?
/usage works from the session history on this machine, and the docs call its figures approximate. Your limits are shared with claude.ai, Claude Desktop and your other computers, whose usage /usage doesn't see.
Do subagents count toward my limit?
Yes. Subagents and agent teammates draw on the same session and weekly limits as the rest of Claude Code. Anthropic says agent teams use approximately 7x more tokens than standard sessions when teammates run in plan mode.
How do I reduce Claude Code's token usage?
Keep sessions focused and start a fresh one for a new task, avoid long idle breaks in a session you'll resume, and check the subagent and MCP rows of /usage for tools you don't need. Anthropic's cost guide has more detail.
Sources
- Commands, Claude Code docs, checked 24 September 2026
- Manage costs effectively, Claude Code docs, checked 24 September 2026
- Prompt caching: cache lifetime and choosing the TTL yourself, Claude Code docs, checked 24 September 2026
- How do usage and length limits work?, Anthropic help center, checked 24 September 2026
- UsageMeter changelog, UsageMeter Link 0.5.0
What changed
- 24 September 2026: first version.