Tamp Compression Proxy Reduces Coding Agent Token Usage by 52%, Works Across Claude, Gemini, and OpenAI Models
Key Takeaways
- ▸Tamp achieves 52.6% reduction in input tokens for coding agents with zero code changes required
- ▸Supports multiple AI providers and agent types through a universal compression proxy architecture
- ▸Uses multi-stage compression pipeline including JSON minification, TOON columnar encoding, and neural text compression
Summary
Sliday has released Tamp, an open-source compression proxy designed to reduce token consumption for AI coding agents by 52.6% without requiring code changes. The tool acts as a transparent intermediary that compresses API requests before forwarding them to Claude, Gemini, Aider, Cursor, Cline, Windsurf, and other OpenAI-compatible agents. Tamp employs multiple compression techniques including JSON minification, columnar encoding, line-number stripping, and neural text compression via LLMLingua-2 to intelligently reduce input token size while preserving code quality and functionality.
The proxy supports multiple API formats including Anthropic Messages, OpenAI Chat Completions, and Google Gemini APIs, with easy installation via npm, git clone, or a one-line installer script. Users simply set an environment variable pointing their agent to the local proxy (http://localhost:7778) and Tamp handles compression silently in the background. The compression benefits compound across multiple turns since the full conversation history is re-sent with each API call, and an in-memory cache prevents duplicate compression of identical content.
- Fully open-source with easy installation and transparent setup requiring only environment variable configuration
Editorial Opinion
Tamp addresses a real pain point for developers using AI coding agents — token costs and API rate limits. The 52% reduction in token usage is substantial and could meaningfully lower costs for teams relying on these tools. The zero-code-change approach and transparent proxy architecture make adoption frictionless, though the effectiveness of the compression will ultimately depend on the types of tool outputs each coding task generates.
