gnaw

A terminal UI for choosing exactly what a model sees. Browse the tree, toggle files and directories in or out, and watch the token total update as you go — a context window is a budget, and the tool should show you spending it.

Why it exists

Agents pick their own context now, and most of the time that's the right default. But not always: code review across a whole subsystem, a migration plan, a raw API call where nobody is doing retrieval for you. In those cases I want to be the retrieval step — deterministic, inspectable, and certain that .env.production didn't tag along.

Three requirements: fast, interactive with visible token costs, secret-scanned. No existing tool had all three. The fast ones are fire-and-forget; the interactive one doesn't show token costs. So gnaw is a fork of code2prompt with live token accounting, a parallel extraction engine, and gitleaks wired in as the default scanner.

Numbers

Packing the Next.js monorepo (26,706 files), Docker-pinned at 8 cores, hyperfine over 10 runs:

The interesting number isn't the headline multiple — it's the gap between those first two rows. A full secret scan costs about 200 ms. Whatever era "scanning is too slow to leave on by default" belonged to, it's over.

Against repomix-rs with both tools scanning it's a dead heat, 1.01 ± 0.08, at the cost of 166 MB more peak RSS for the larger ruleset. The full methodology, caveats, and the part where I misread my own benchmark by 5x are in the write-up.

Next

An MCP server, so agents can call gnaw as a token-budgeted, secret-scanned context extractor rather than this staying a manual-only tool.

Repo on GitHub — quickstart in the README, benchmark script and Dockerfile beside it.