Short answer
Only if the tool is in front of it every session: agents are trained on grep and fall back to it. With Pixel 0.5.0 wired by pixel install, Claude Code called Pixel in 11 runs out of 11, 6 to 17 times each. An earlier release, given its protocol alone with no hooks, was barely called at all.
Why a tool goes unused
An agent is trained on the tools every repository has: search, read, the shell. A new tool it is told about once, in a file it may or may not load, competes with the habits it was trained on, and the habits often win. Listing a command in a README is not enough.
What made the difference
Pixel does not wait for the agent to discover it. pixel install puts its protocol into every session of the agents it supports, and on Claude Code it does so through lifecycle hooks, so the guidance is there on the first turn of every session. pixel install --repo adds a guard to one repository that flags untargeted reads of large files; it advises and never blocks.
The difference shows between releases: the one behind the recorded demo was called in every run, where an earlier release, given its protocol alone, was barely called.
Check it on your machine
pixel doctor reports whether the hooks and the protocol are in place for each agent, and pixel action-log reads the local log of every Pixel command that ran, so you see whether your agent called it. The page for your agent lists what pixel install writes for it.
The figures
| Runs in which the agent called Pixel | 11 of 11 |
|---|---|
| Pixel calls per run | 6 to 17 |
| Earlier release, protocol only, no hooks | Barely called; tasks still scoped 29% faster |
Every figure above is on the benchmarks page, with its sample size and its method: On whole agent tasks.
Where Pixel does not win
One agent, Claude Code, on one task: how often another agent calls Pixel is not measured. The per-repository guard advises and never blocks, so an agent can still read a whole file. Part of the gain is the protocol's guidance rather than its answers: the earlier release that was barely called still scoped tasks 29% faster.