How to check Claude Code token usage: 4 ways to see what your agents spend

If you run Claude Code for more than a few minutes a day, tokens are your real currency. They decide when you hit a rate limit, how fast you burn through a subscription, and whether a long agent run was worth it. The problem is that Claude Code does not shout the number at you. You have to go and look.

Here are four ways to check Claude Code token usage, from the quickest built-in command to a live meter that shows every agent in real time. Pick the one that fits how deep you need to go.

1. The fastest way: the /cost command

Inside any Claude Code session, type:

/cost

Claude Code prints the total cost and token count for the current session, plus how long it has been running. It is the fastest answer to "what has this session spent so far", and it needs nothing installed.

The catch: /cost is scoped to the session you are in. It does not tell you what you spent yesterday, what a different project cost, or how a specific agent compares to another. For a single quick check it is perfect. For anything historical, keep reading.

2. Read the source: the session transcripts

Claude Code writes every session to disk as a JSONL transcript. On macOS and Linux you will find them here:

~/.claude/projects/<encoded-project-path>/<session-id>.jsonl

Each line is one message, and the assistant messages carry a usage object with the exact counts: input_tokens, output_tokens, and the cache fields (cache_creation_input_tokens, cache_read_input_tokens). Cache reads are the interesting part: they are much cheaper than fresh input, so a session with a high cache hit rate costs far less than the raw token count suggests.

This is the ground truth. Every other tool, including the two below, reads these same files. If you want to script your own report, jq over the JSONL will get you there.

3. A dedicated CLI: ccusage

If you would rather not write your own parser, the community ccusage tool reads those transcripts and rolls them up into daily, monthly, and per-session totals:

npx ccusage@latest

It is a good middle ground: more history than /cost, less work than hand-rolling jq. It stays in the terminal, though, so it answers "what did I spend" but not "which agent, right now, is spending it".

4. A live meter: per-session usage in AgentsRoom

The three methods above all share one blind spot: they are after-the-fact and single-stream. The moment you run several agents in parallel, you want to see usage per agent, live, without typing a command in each terminal.

That is what the Claude Code token usage meter in AgentsRoom does. It reads the same JSONL transcripts, per session, and surfaces:

• input, output and cache tokens per agent, updating as the session runs
• the cache hit rate, so you can see when context is being reused instead of re-sent
• a warning before a session drifts toward a rate limit

Because it attributes usage by session, it stays correct even when you run multiple Claude Code accounts side by side: each account keeps its own totals, which makes per-client rebilling straightforward.

Which one should you use

• Just need this session's number? /cost.
• Want to script a custom report? Read the JSONL transcripts directly.
• Want daily and monthly totals in the terminal? ccusage.
• Running several agents and want to watch usage live, per agent and per account? Use the meter in AgentsRoom.

Checking usage is step one. If the numbers make you wince, the next move is bringing them down: see how to reduce Claude Code token costs for the levers that actually matter, starting with cache reuse.