Short answer
Cut the reads that bring whole files into context: let the agent ask an index what a file contains, who calls a function and which files a task touches. On one scoping task, 11 runs per side, Claude Code with Pixel read 38% fewer tokens (median 10,655 against 17,268) and cost 30% less. pixel token-savings then reports what it spared on your own sessions.
What the agent spends tokens on
Before Claude Code edits anything, it finds out where to work: it searches, opens files, follows imports and opens more files. Each file it opens stays in the context and is read again, as input, on every later turn. The fewer whole files enter, the less every turn costs.
What changes with an index
With Pixel wired in, the agent asks instead of opening: pixel scope-task for the files a task touches, pixel list-signatures for what a file contains, pixel who-calls for the callers of a function. Each answer is fitted to a budget and says whether it is complete. On the demo’s scoping task both sides named the file to change among their first two in every run, so the saving did not come from missing it.
pixel install wires it into Claude Code (lifecycle hooks in ~/.claude/settings.json), and pixel install --repo adds a guard in the repository that flags untargeted reads of large files. The Claude Code page lists every file it writes and how to remove it.
Measure it on your sessions
pixel token-savings reads the local action log and reports what Pixel’s answers spared, each part labelled measured or estimated. Run it after a few days of normal work; it tells you more about your code than any benchmark of ours.
The figures
| Tokens read into context, median | 17,268 without Pixel, 10,655 with it (−38%) |
|---|---|
| API cost, median | $0.394 without, $0.274 with (−30%) |
| Wall time, median | 86.5 s without, 60.7 s with (−30%) |
| Sample | 11 runs per side, Claude Sonnet 5, Pixel 0.5.0 |
pixel token-savings on the maintainer's sessions | 41 to 83% across 798 operations |
Every figure above is on the benchmarks page, with its sample size and its method: On whole agent tasks, Reading code.
Where Pixel does not win
One task in one repository, so read it as a direction: the spread is wide on both sides, 57 to 148 s without Pixel and 33 to 207 s with it. In the August runs Pixel was about 1.5 seconds slower on a single lookup, the cost of reading its guidance, and it is still slower at recovering deleted code from history. Other levers (a shorter CLAUDE.md, fewer MCP servers, compacting sooner) are not measured here.