Pxpipe is an open-source local proxy that converts long Claude Code inputs into compact PNG images to reduce token costs, effectively compressing bulky content before it reaches the model.

By rendering system prompts, tool docs, and older history as images, pxpipe dramatically lowers input tokens while leaving model outputs unchanged.

Benchmarks show substantial end-to-end token savings, typically in the 59% to 74% range, though results vary with workload, document density, and the model used.

Imaging is lossy and not byte-exact; some content (IDs, hashes, exact numbers) must remain text to avoid silent errors, and not all models or inputs benefit equally.

Benchmarks and findings are hosted in the pxpipe repository, with The Decoder cited as a source for AI news and analysis.

Default model support includes Claude Fable 5 and GPT 5.6, with additional models like Opus 4.7/4.8 and GPT 5.5 explored in benchmarks; accuracy varies and features can be toggled manually.

Historical context notes that compressing text as images has precedents, such as Deepseek’s OCR-based approach achieving high reduction while preserving most information.

Practical demonstrations compare pxpipe-enabled runs against baselines, showing significant token reductions and lower perceived costs in measured traces.

Quick-start guidance is provided, including model configuration options, a no-proxy rendering path for PNGs, and a live dashboard to monitor token savings and per-session stats.

Open-source with MIT license, containerized deployment, and guidance to reproduce results from local event logs, with caveats about relying on imaging for exact recall.

In practice, pxpipe renders bulky static content like system prompts, tool docs, and older chat history as images, while recent messages and model outputs remain in text.

The proxy intercepts /v1/messages, re-renders bulk history into image blocks, preserves prompt caching, and logs per-request events for token accounting and auditing.