pxpipe iconpxpipe

open-source Free Star6k

Local proxy that cuts Claude Code token bills 59–70% by re-rendering bulky context — system prompts, tool docs, and collapsed history — as compact PNGs, exploiting the fixed image-token cost of Fable 5's vision channel.

typescript api available self hosted cost reduction
59–70% End-to-End Token Savings
~3.1× Chars per Image-Token vs Text
4.8K+ GitHub Stars

Overview

pxpipe is a local MITM proxy that cuts Claude Code's input token bill by rendering bulky context — system prompts, tool docs, and older collapsed history — as compact PNGs before each request leaves your machine. An image's token cost is fixed by its pixel dimensions, not by how much text is inside it. Dense content (code, JSON, tool output) packs ~3.1 chars per image-token vs ~1 char per text-token on real Claude Code traffic. The proxy intercepts /v1/messages, rewrites eligible bulk into image blocks behind a per-request profitability gate, and forwards — responses stream normally. A live dashboard at 127.0.0.1:47821 shows tokens saved, text→image conversions side-by-side, and a kill switch. Recent turns always stay text; only the static prefix and older bulk history are imaged.

The Verdict

Who Should Use pxpipe?

Best For

  • Claude Code users on Fable 5 with token-dense workloads (code, JSON, logs)
  • Teams running long sessions with large system prompts and tool docs
  • Developers who want measurable, per-request cost accountability via events log
  • Projects where 59–70% end-to-end token savings justify some lossiness on imaged content

Not Ideal For

  • Workflows requiring byte-exact recall of hex strings, IDs, or secrets from context (lossy by design)
  • Opus 4.8 users — misread rate ~7% on rendered pages, opt-in only
  • Sparse-prose workloads (~3.5 chars/token) where text wins on cost
  • Teams needing zero added request latency (PNG encoding adds time before forwarding)

What's Great

  • 30-second setup: npx pxpipe-proxy + one env var points Claude Code at it
  • 59–70% end-to-end bill reduction measured across all requests, not just compressed ones
  • Savings measured honestly: free count_tokens probe on each original request, compared against actually-billed usage, no double-counting of cache discounts
  • Fable 5 reads imaged context at 100/100 on novel arithmetic and 98/98 on gist recall benchmarks
  • SWE-bench Lite pilot: 10/10 both arms at −65% request size
  • Profitability gate — sparse prose stays text, images only when math wins
  • MIT licensed, TypeScript, usable as a library without the proxy
GitHub README · FINDINGS.md

Watch Out For

  • Lossy by design: verbatim 12-char hex recall drops to 13/15 on Fable 5, 0/15 on Opus — silent confabulation, not hard errors
  • PNG encoding adds latency to large requests before they leave your machine
  • Opus 4.8 and GPT 5.5 are deliberately opt-in — both read imaged content measurably worse
  • Workload-dependent: wins on token-dense content, loses money on sparse prose
  • Rendering research paused as of 2026-07-05 — verbatim misreads are capacity-bound, not solvable by layout tweaks
GitHub README — The honest part & Limitations

Pricing

Open Source
Free
MIT license — proxy, dashboard, library API, and all benchmarks included
View all features & details

What Gets Compressed

  • Large tool_result bodies (file reads, command output, logs) above ~6k chars of token-dense content
  • Older collapsed history: turns behind the live tail re-rendered as image pages
  • Static system prompt + tool docs slab
  • Everything else passes through byte-identical (messages, recent turns, model output)

Benchmarks (Fable 5)

  • Novel arithmetic (N=100): 100% accuracy, −38% tokens
  • Gist recall with distractors (15k–45k char sessions, N=98/arm): 98/98 both arms
  • State tracking (N=18/arm): 18/18 both arms
  • Confabulation on never-stated facts (N=16/arm): 0/16 both arms
  • SWE-bench Pro: 14/19 ON vs 15/19 OFF at −60%, 18/19 verdicts agree

Model Scope

  • Default: claude-fable-5 and gpt-5.6
  • Opus 4.8 and GPT 5.5: opt-in via PXPIPE_MODELS or dashboard (measurably worse at reading renders)
  • PXPIPE_MODELS=off disables imaging entirely
  • Subagents on non-allowlisted models pass through as text

Technical Details

  • 1928px-wide columns, ~92,000 chars/page, ≈4,761 vision tokens per image
  • Prompt caching preserved: static prefix splice kept cache-friendly
  • Events log at ~/.pxpipe/events.jsonl — per-request counterfactual and actual tokens
  • Library mode: renderTextToImages / transformAnthropicMessages for direct integration
  • Pure-JS runtime (Node and edge/Workers); @napi-rs/canvas build-time only

How It Compares

Feature pxpipe Token Optimizer Context Mode Manual Compaction
Approach Image compression proxy Context pruning plugin Selective context Manual /compact
Token Savings 59–70% end-to-end Workload-dependent Workload-dependent Session-dependent
Setup 30 seconds, one command Plugin install Toggle flag Manual trigger
Lossiness Imaged content (manageable) None None None
Accuracy Measured Per-request counterfactual Aggregate Estimate None
Model Scope Fable 5, GPT 5.6 (default) All All All

User Reviews

Loading reviews...