Claude Code usage: how to track it and make it last

Q: How do I check Claude Code usage?

Run /usage inside Claude Code. It shows your plan limit bars and a breakdown of what used them, with a 24-hour and 7-day view. /status shows your model and context instead, not plan limits.

Last checked: September 15, 2026.

Run /usage inside Claude Code to see your plan limit bars and what's using them. Claude Code draws from the same session and weekly limits as claude.ai and the desktop app, so a heavy coding afternoon is why chat runs out later.

The single biggest lever isn't which model you pick. It's context: every turn re-reads the conversation and the files it has opened, so a long session costs progressively more per turn.

Seeing what you've used

/usage in Claude Code shows your plan limits and a breakdown of what consumed them, including skills, subagents and MCP servers, with a 24-hour and 7-day toggle. /cost and /tokens reach the same place.
Settings → Usage on claude.ai shows the same plan limits from the other side: current session, weekly across all models, and per-model weekly bars.
/status is a different thing: model, context and configuration, not plan limits.
/usage-credits opens the settings page where usage credits are managed.

The /usage breakdown counts the machine you're on. If you run Claude Code on a laptop and a desktop, each one only knows its own share, while the plan bars cover everything.

Where the usage actually goes

Four things dominate, in roughly this order:

Session length. Context accumulates: the conversation, every file read, every tool result. Turn fifty in a long session can cost many times turn five, which is why limits so often arrive late in an afternoon that felt fine.
Large file reads. Reading a whole large file into context charges for it on that turn, and again on the turns after it.
Long tool chains and agents. Every tool result joins the context. Subagents and MCP servers show up in the /usage breakdown for exactly this reason.
Model choice. Real, but usually smaller than the first three. Switching models also doesn't bypass a limit: Anthropic's documentation says so plainly.

Watch the meter without leaving your editor.

AI Limit Coach keeps your session and weekly usage visible in Claude in the browser, warns when you're on pace to run out before the reset, and notifies you when it resets, so a long Claude Code afternoon doesn't end in a surprise. Free.

Get the extension

Habits that stretch a five-hour window

/clear when you change task. Carrying yesterday's debugging into today's feature work costs you on every single turn.
/compact before the session gets huge, rather than after it's already slow and expensive.
Keep CLAUDE.md short and specific. It's re-read constantly, so every stale paragraph is a recurring charge.
Ask for targeted reads. "Read the failing test and the function it calls" costs a fraction of "read the whole module".
Batch related asks into one turn instead of ten small ones, each of which re-reads everything.
Check the meter before a big job. Starting a large refactor with 15% of the window left wastes both the window and the refactor.

Newer versions of Claude Code make the end of a window less disruptive: from v2.1.234, an open session can wait and continue on its own once the limit resets.

Measuring it yourself, and the trap in doing so

Claude Code writes session transcripts locally under ~/.claude/projects, and they contain token counts. Several open-source dashboards read them to estimate spend, and you can too.

One warning from doing exactly that: resuming or branching a session copies earlier messages into the new transcript file. Adding up raw records double-counts them badly. On one machine here, raw totals came out 94% too high: 202,000 records for 95,000 real API calls. Every record carries a request id, so count each id once across all files and the number lands where it should.

Two more caveats worth knowing before you trust a local total: it only sees the machine it runs on, and it only sees Claude Code, not claude.ai or the phone. For limits, the plan bars remain the source of truth.

Questions people ask

How do I check Claude Code usage?

Run /usage inside Claude Code. It shows your plan limit bars and a breakdown of what used them, with a 24-hour and 7-day view. /status shows your model and context instead, not plan limits.

Does Claude Code share limits with claude.ai?

Yes. On Pro and Max, Claude Code, claude.ai and the desktop app all draw from the same session and weekly limits, so a heavy coding afternoon reduces what's left for chat.

Why does Claude Code use so much of my limit?

Context size. Every turn re-reads the conversation, the files it has opened and the tool output it has collected, so cost grows as a session goes on. Long sessions, large files and long tool chains are the expensive parts.

Do local usage totals match my plan usage?

Not exactly. Local transcripts cover only that machine and only Claude Code, and resumed or branched sessions duplicate earlier messages in the files. Count each API call once, by its request id, or totals come out far too high.

Is Claude Code the reason you keep running out?

Measure a week of real percentages, then let the plan calculator tell you whether a bigger plan is worth it or your sessions just need trimming. The extension collects those percentages for you.