In 58 days of my Claude Code sessions, 49% of tool calls ran shell commands. The call counts, cost at API list prices and cache figures from one machine.
For 58 active days I had been running Claude Code, yet I could not have told you what its tool calls were doing. The evidence was already in transcript files on my disk. I wrote Caprock to read them and found that 49% of those calls ran shell commands.
$10,513 over 58 days
$10,513 at Anthropic's list prices, 58 days
$181 a day, on the days I worked
$200 what the plan costs a month
That is what my sessions would have cost at Anthropic's list prices. I pay $200 a month for a plan, so it is not a saving — it is what the same work would cost billed per token.
Nearly half the tool calls ran Bash
| Share | ||
|---|---|---|
| Bash | 49% | 40,213 calls |
| Read | 14% | 11,681 calls |
| Edit | 9% | 7,862 calls |
| WebFetch | 7% | 6,052 calls |
| WebSearch | 4% | 3,525 calls |
Share of all tool calls, 58 days on one machine.
Nearly half my tool calls ran commands: tests, builds, searches and git. Edit and Write together accounted for under 12%.
I had it backwards: I thought of Claude Code as something that writes code and occasionally checks its work. On my machine it reached for the toolchain far more often than for Edit or Write. These are call counts, not shares of cost: one call can consume far more tokens than another.
Command output also becomes part of the context that later turns read. The counts alone cannot tell me how much of the token cost came from that output.
A 99% cache hit rate cut input cost by 89%
99% of input read from cache
89% of input cost cut by it
The 99% hit rate is not a trick I found; Claude Code caches prompts by default. It cut input cost by 89% in this sample. Cached reads still cost money, and a longer conversation keeps getting read.
Read your own record
Do not change your workflow because Bash topped my chart. Count your own calls, then look at tokens and prices. The first tells you what the agent did; the second tells you what it cost.
Caprock is a free, open-source desktop app that keeps the ledger of every coding agent on your machine: who is waiting on you, who is stuck, how close the limit is, and what each repository and pull request cost at list price. It reads your own transcripts, local by default, so the costs and tool-call counts you see are your own.
Sources
- Anthropic API pricing — the list prices used
- The same machine's later reading, with the method
- Claude Code: prompt caching — handled automatically
Figures read on 2026-08-25 by Caprock v0.23.1 on one machine, mine. A reading, not a live feed; yours will differ.
First published on caprock.dev. Caprock is the free, open-source app I build to run and watch Claude Code, Codex and other coding agents on my own machine: github.com/dspv/caprock.
Top comments (0)