We Measured It: A Code Map Cuts ~24% of Your AI's Coding Tokens

Every time you ask an AI assistant a question about your codebase, it starts by re-reading your codebase. It greps, it opens files, it re-derives the structure it could have been told — and that reading is most of the token bill.

We built a fix: a deterministic, committable code map that your assistant queries instead of re-reading files. Then we did the harder part — we measured whether it actually saves tokens, in the open, without shaping the test to flatter the result.

We pre-registered the test so we couldn't cherry-pick

The easy way to produce a good benchmark number is to choose the tasks after you know which ones your tool wins. So we didn't. Before running anything, we wrote down the task mix and the pass/fail bar and committed to them — anchored to how real issues actually read.

Real GitHub issues are symptom-described: "the total is wrong after a promotion", "refunds omit the tax" — not "change function X". So the test is mostly symptom tasks, with a small floor of trivially greppable ones kept in on purpose.

Then we ran the same coding tasks twice with the same model — once with ordinary file tools, once with the Vibgrate code map — and counted only the tasks both runs solved. Fewer tokens for the same result, never cheaper because it gave up.

The result: ~24% fewer tokens, same results

On a ~300-file codebase, the assistant with the map spent about 24% fewer tokens with the same verified results. On the discovery-heavy tasks — find the right code from a symptom, then fix it — the median saving was +33.8% across three runs.

We publish the floor, too

On tiny repositories the map saves nothing — grep finds the code instantly, and the map's fixed cost shows up as a loss. We publish that part of the curve as plainly as the win, because a benchmark that only shows its best number is not one you should trust. The claim is the curve, not a single flattering figure.

Try it — free and offline

The code map is free, runs locally, and never sends your code anywhere. It speaks MCP, so it works with Claude, Codex, Cursor, and any MCP-capable assistant.

vg install

That wires the map into your assistant. See the full task-by-task numbers, floors included, on the token-savings benchmark page.