distil

claude mcp add distil -- npx -y @munhq/distil

No account, no API key, nothing to configure. The package is a small wrapper that fetches the binary for your platform and verifies it against the published checksums; install.sh and a prebuilt binary remain for anyone without Node.

---

Cache writes are 2% of your tokens and 28% of your bill

That is not a claim, it is a measurement: 13,814 real agent sessions, 410,742 assistant turns, re-run on 2026-08-29. Every unique token in them was billed 550 times, because a request resends the whole history.

Which means the obvious move — compress the history — is usually the wrong one. Editing history invalidates the cached prefix from the edit onwards and converts reads at 0.1x into writes at 1.25x or 2.0x. You cannot compress your way out of context cost. You can only decline to admit tokens.

Every tool in this space publishes a savings percentage measured on its own fixtures. What none publishes is the denominator: what share of a real session it is allowed to touch, and what that share costs once prompt-cache pricing is applied. distil measures both on transcripts an agent actually wrote.

What is prior art, and what is not. The cache arithmetic below is not a discovery. Anthropic's context editing docs state that clearing tool results invalidates the cached prefix, and ship clear_at_least so a clear only fires when it is large enough to pay for that. The break-even rule is published too: on a 5-minute cache, cleared tokens times requests-before-the-next-clear must exceed 11.5 times the tokens you keep. The table in this README reproduces that rule exactly — it was derived independently, which is a check on the arithmetic, not a contribution.

The gap is empirical. Every source says to calibrate against your own workload, and none ships a way to do it or publishes what the values turn out to be. That is what this crate is for: measuring the numbers you need in order to choose clear_at_least, or to decide not to clear at all.

The full measurement

Measured 2026-08-29 over 13,814 local Claude Code transcripts — 410,742 assistant turns, 212.7M tokens of unique text. Reproduce it on your own corpus with distil-bench ~/.claude/projects; a corpus grows, so the date matters more than the decimals.

Those 212.7M unique tokens were billed as 116.9 billion input tokens — every token paid for 550 times, because a request resends the whole history.

Price that at real cache multipliers (read 0.1x, write 1.25x for the 5-minute TTL and 2.0x for the 1-hour):

Cache writes are 2% of the tokens and 28% of the bill. Editing history invalidates the cached prefix from the edit onwards, converting reads at 0.1x into writes at 1.25x or 2.0x. So a rewrite must shrink what it invalidates below:

That table is the published break-even rule in another form: at every row, cleared x turns / kept equals 11.5 for the 5-minute tier. Use it to pick a clear_at_least value, and use distil-bench to find the turn count and tail size to put into it — those are workload properties, and they are the part nobody publishes.

Per unit of history, at 10 remaining turns: keeping it costs 1.00, compressing it costs 2.15, and never admitting it costs 0. You cannot compress your way out of context cost. You can only decline to admit tokens.

What that means for using this crate

Layers that do not touch history are on the right side of that arithmetic: CacheAlignLayer (orders content so the stable prefix stays cacheable) and ScratchpadLayer (keeps working state outside the window).

Layers that rewrite history — MaskingLayer, SummarizationLayer, CompactionLayer — cost more than they save in the common case. Reach for them at one boundary only: context overflow, where the alternative is a failed request and cache price stops being the comparison. BudgetLayer exists for exactly that moment.

RegistryLayer and CodeModeLayer predate Anthropic's Tool Search Tool and Programmatic Tool Calling, which do the same jobs natively and better. Prefer the native features.

For clearing old tool results, prefer the provider's clear_tool_uses context editing over MaskingLayer: it runs server-side, it takes clear_at_least, and it is one API parameter against a dependency. Reach for a layer here only when you need behaviour the API does not offer.

Measuring

cargo build --features bench --release

Where tokens are, what they cost, and the break-even table

./target/release/distil-bench ~/.claude/projects --json baseline.json

Sessions that called a given tool, against those that did not

./target/release/distil-bench ~/.claude/projects --split-by-tool mcpcodeindex

Export real traffic so other compressors run on the same input

./target/release/distil-bench ~/.claude/projects --export-sessions ./sessions --min-turns 40

See bench/README.md for the external-tool comparison, the fairness rules, and the two harness mistakes that produced wrong numbers first.

Retention

A saving is only a saving if the model can still answer what the original context could answer.

# No LLM judge: file paths checked against ground truth from the transcript

python bench/artifact_retention.py ./sessions 12

LLM-graded probes (recall / artifact / continuation / decision)

cargo build --features probe --release ./target/release/distil-probe --probes 6 --model qwen2.5:3b

The probe taxonomy is Factory.ai's; their write-up defines it and ships no harness. The judge is a Completer, never a Summarizer — a summarizer may impose summarization framing, which rewrites both the probe format and the grading instruction.

Using it as a library

use distil::{CacheAlignLayer, Ctx, EstimateCounter, Pipeline};

let pipeline = Pipeline::builder() .counter(EstimateCounter) .layer(CacheAlignLayer::generic()) .build();

let mut ctx = Ctx::new(messages, tools, turn); let result = pipeline.optimize(&mut ctx); println!("{result}");

This example is kept compilable as examples/readme_quickstart.rs — run it with cargo run --example readme_quickstart.

Every layer implements Layer and reports tokens_before, tokens_after and a detail line, so each one can be measured on its own.

Features

Install

./install.sh                       # binaries, the skill, and the MCP server

/plugin marketplace add munhq/distil /plugin install distil # Claude Code: skill and server in one step

install.sh installs both binaries, drops the skill into every Claude home it finds, and registers the MCP server at user scope. When the plugin is already installed it installs the binary only, since the plugin declares the server and ships the skill itself.

The plugin launches the server with npx -y @munhq/distil, so it needs Node. It cannot use a plugin-relative path: Claude Code expands ${CLAUDE_PLUGIN_ROOT} and nothing else does, so a plugin declaring one hands every other client a literal path that does not exist. install.sh and the prebuilt binaries remain for anyone without Node.

Platform support

The release publishes six targets and plugin/test_platform.sh holds both the installer and the plugin launcher to that matrix, so an asset name and the name asked for cannot drift apart. install.sh and the launcher are bash scripts, so on Windows they need a shell — cmd and PowerShell cannot run them. Linux binaries are static musl builds, so they do not need a matching glibc.

Caveats

The corpus is one developer's machine. The ratios are the finding; the absolute totals are personal. Counts use cl100k_base, which approximates Claude's tokenizer within a few percent. The break-even model assumes a single cache breakpoint, so a rewrite confined to the tail costs less than the table shows — that refines it, it does not reverse it.

Contributing

Build and test instructions, the rules a benchmark change has to follow, and what a pull request needs before review: CONTRIBUTING.md. Report a vulnerability privately — SECURITY.md.

License

MIT. See LICENSE.