Developer tool · v0.1 · MIT
whittle collects the techniques that cut a coding agent’s token bill and turns each one into a lever you switch on per repository. It wires itself into Claude Code and Codex from your home directory and writes nothing into the repositories it serves.
$ whittle install
claude-code: written ~/.claude/settings.json (5 appended)
codex: not installed (~/.codex does not exist); skipped
nothing is active yet: run `whittle enable` inside a repository to opt it in
$ whittle enable
enabled ~/code/shop (repo a5106c195910)
shell-output: off
prefix-audit: on
$ whittle prefix-audit
TOKENS CHARS KIND SOURCE
~644 2506 claude-md ~/code/shop/CLAUDE.md
total ≈ 644 tokens across 1 components
not measured: Claude Code's own system prompt and built-in
tool schemas; MCP server instructions and tool schemas; …
$ git status --short
$ ~ and
a few lines are trimmed. The empty git status is the point.
Where the tokens go
Every turn re-sends a fixed prefix, tool schemas and the history so far, on top of tool output, file reads and the model’s own reply. Each lever targets one layer, so you can tell what changed when you switch it.
a shipped lever covers it a lever is planned
Levers
whittle install only wires the hook. Nothing runs in a repository until you
enable it there, and each lever can be switched on its own.
shell-outputopt-inL3 · tool output · hook
Rewrites allowlisted Bash commands through rtk, which filters their output before the agent reads it. Enforced in Claude Code; Codex gets the hook but not the rewrite.
Needs the rtk binary, which you install yourself. A published paired run measured
+7.6% cost with it, so whittle shows that figure, labelled as someone else’s.
prefix-auditon when enabledL1 + L2 · fixed prefix · report
Ranks what rides along on every turn, memory files, rules and skill descriptions, in approximate tokens, and lists what it could not measure. It reports only and injects nothing.
Counts use the o200k tokenizer as an estimate; Claude’s own tokenizer is not public.
retrievalnot yet shippedL4 · code retrieval
Will answer “where is this, who calls it” from a code graph instead of whole-file reads, through graft.
Three rules
Hooks go into your user-level harness config; state and caches live under the XDG directories.
whittle footprint lists every path it has ever written, and the test suite fails if
a stray file appears in a repo.
No telemetry, no hosted service, no network calls at runtime and no install-time scripts. The hook fails open and always exits 0, with a time cap per lever.
A lever shows what was measured on your machine or says “unmeasured”. Figures from elsewhere are printed as external evidence with their source.
Commands
install / uninstall-y, and reports whether each config came back byte-identical to its backup.enable / disableleversstatus--json.doctorfootprintprefix-auditInstall
whittle is not on the npm registry yet. Install the packed release from this site, or build it from source. Either way the hook runs under Node, which starts fastest.
npm install -g https://whittle.aylith.com/downloads/aylith-whittle-0.1.0.tgz
whittle install # wire each detected harness
whittle enable # opt the current repository ingit clone https://github.com/aylith-labs/whittle.git
cd whittle && bun install && bun run build
npm install -g .
whittle install && whittle enable
Checksums: SHA256SUMS. Undo everything with
whittle uninstall -y.
The shipped binary is plain JavaScript with two dependencies.
install wires whichever it finds in your home directory and skips the rest.
whittle never installs it. whittle doctor tells you if it is missing.