Cut your Token Usage by 41%

live · session avg

The Distillery sits between your AI coding tool and the Anthropic API, where it deduplicates repeated context and distils large tool outputs without changing your workflow and without your context leaving your machine. Works with Claude Code, Cline, Goose, Aider, Zed, and any OpenAI-compatible tool.

Zero-config. No prompt rewriting. No workflow changes.

$npm install -g thedistilleryCOPY

Create free account See pricing

no card required · cancel anytime

177,000,000tokens saved · and counting

41%avg reduction433.0Mprocessed1,400sessions

Three steps. One terminal.

~/ · install

runs in ≈4s · no signup required

Production-grade optimization.

Four techniques, one smaller bill. Toggle each to watch a real session’s token count drop.

Tokens per session ↘ (Lower is better)

Reduction techniques

01Context Deduplication−25%Hash tool_result outputs across the session. First copy stays verbatim; repeats become a short marker.

02Tool Output Distillation−15%Compress large read_file and bash outputs before they re-enter the next turn.

03Prompt Cachingup to ~10%Place cache breakpoints automatically. Stop paying full rate for the same prefix. Savings vary by session; up to ~10% on cache-friendly workloads.

04Attention-Aware Optimization−8%Compress the conversation’s low attention middle harder. The head and last 3 turns stay protected. Plus tool call supersession when later calls repeat the same input.

Without DistilleryPOST /v1/messages · raw payload

system prompt (2.1k chars)2,104 tok

user: "Refactor middleware to..."312 tok

tool: read_file(auth.ts) ×310,200 tok

tool: bash(npm test) — 1,840 lines8,420 tok

+ 14 earlier middle-zone turns221,474 tok

assistant: next reply4,800 tok

total247,310 tok · $3.29

With DistilleryPOST /v1/messages · context deduplication

system prompt (2.1k chars)2,104 tok

user: "Refactor middleware to..."312 tok

tool: read_file(auth.ts) ×3×3 → ×1 ref3,400 tok

tool: bash(npm test) — 1,840 lines8,420 tok

+ 14 earlier middle-zone turnsduplicate outputs marker’d166,447 tok

assistant: next reply4,800 tok

total−25% · saved $0.82185,483 tok · $2.47

One choice, Two bills.

Option A · do nothing

Keep paying for duplicate context.

Your sessions re-send the same files, the same tool outputs, the same context, turn after turn. Anthropic bills every byte. You never see it itemized.

tokens sent247,310

duplicates~57,080

rate$0.0000133

this session$3.28

Option B · install

Let the proxy delete the waste.

Distillery sits between your CLI and Anthropic, dedupes repeated context, and distils tool outputs before the bytes ever leave. Same agent. Smaller invoice.

$npm install -g thedistilleryCOPY

tokens sent190,230

savings−23%

rate$0.0000133

same session$2.53

Cut my Claude Code bill pricing

Flat EUR 9/mo. Keep 100% of what you save.