#9393rd of 98 in Skills, plugins and rules
pxpipe
Local proxy that renders bulky LLM request context as PNGs to cut input tokens
- Stars
- 7.5k
- License
- MIT
- Last commit
- Oct 2026
- Last release
- Sep 2026
- Language
- TypeScript
Missed the latest GitHub check; these facts may be older.
Overview
pxpipe sits between Claude Code (or Codex, cursor-agent, and other clients) and the provider API. It rewrites the system prompt, tool docs and older history into dense PNG pages plus a bounded factsheet before the request leaves the machine. It handles Anthropic Messages, OpenAI Responses/Chat Completions and Google generateContent, and logs measured savings against a count_tokens counterfactual.
Who it is for: Heavy Claude Code users trying to cut input-token spend
Strengths
- Savings measured per request against a count_tokens probe, logged to ~/.pxpipe/events.jsonl
- Reported ~59-70% lower end-to-end bill on Claude Code traffic; SWE-bench Lite 10/10 both arms
- Profitability gate leaves sparse prose and recent turns as text
- Also offers offline PNG export, a library API and a dashboard on 127.0.0.1:47821
Weaknesses
- Lossy: exact 12-char hex recall 13/15 on Fable 5, 0/15 on Sol; misses are silent
- No verbatim-risk guard yet; IDs and hashes must stay text
- Default model scope is limited: Fable 5, Opus 5.5 and Gemini; others opt-in
- SWE-bench Pro showed 14/19 vs 15/19 text; small samples, and the README says savings depend on workload
What it needs
- no GPU
- Docker + Compose
- Needs Node.js, Anthropic, OpenAI-compatible or Google API access
- Models: claude-fable-5, claude-opus-5-5, claude-opus-5, Gemini 3.6/3.7 Flash, gpt-5.6-sol (opt-in)
- port 47821
Also in Skills, plugins and rules
See all 98| Rank | Project | Score |
|---|---|---|
| 1 | skillsAgent skills for planning, TDD, debugging and code review | 94 out of 100 |
| 2 | eccAgents, skills, hooks and rules bundle for Claude Code, Codex and other harnesses | 94 out of 100 |
| 3 | cavemanCuts token use in coding agents with terse replies and compressed inputs | 94 out of 100 |
| #4 | impeccableDesign skill with 24 commands and a detector for AI-generated frontends | 93 out of 100 |
| #5 | ponytailSkill prompt that makes coding agents write less code and add tests | 92 out of 100 |