Each question here has a figure we measured ourselves: the answer comes first, then the detail, the figure with the section of the benchmarks page it comes from, and where Pixel does not win. A question without a figure of ours does not get a page.
Why does my coding agent read whole files?
Its read tool returns files, not answers; an index answers what a file contains in a fraction of the tokens.
MCP server or CLI: what does an agent tool cost per turn?
An MCP server costs its tool schemas on every turn, a CLI the prompt that teaches it; the size decides, not the protocol.
How do I reduce Claude Code's token usage?
Cut the whole-file reads: on one scoping task, 11 runs per side, Claude Code read 38% fewer tokens with Pixel.
Grep or a code index for a coding agent?
Grep finds the line; the whole-file read after it is the cost an index removes.
Will my agent actually use the tool I give it?
Only if the tool is in every session: with Pixel 0.5.0, Claude Code called it in 11 runs out of 11, 6 to 17 times each.
How do I know what a change will break before the agent edits?
Ask for the blast radius first: pixel impact found 0.86 of the callers on 29 cases, with a warning when an answer may be incomplete.