Squeezr

Squeezr

Token compression proxy for AI coding CLIs. Sits between your CLI and the API, compresses context on the fly — without ever breaking Anthropic's prompt cache — and saves thousands of tokens per session. Real-time dashboard, MCP integration, and an optional AI compression layer powered by Zest, Squeezr's own model.

Supported CLIs & apps

Works with both API keys and subscription plans (OAuth) — Claude Code Max/Pro, OpenAI Plus, etc.

Quick start

Prompt-cache safety (the core design principle)

Anthropic bills cached context at 0.1x — but only if the request prefix arrives byte-for-byte identical between requests. A proxy that mutates history differently on each request silently invalidates that cache and re-bills the full context at full price, costing far more than compression saves.

Squeezr is built around this constraint:

- Deterministic compression is byte-stable — same input always produces the same output (fixed-pressure rules), so the cached prefix never changes → Anthropic's cache keeps hitting.
- Unstable passes never touch the cached prefix — AI compression, cross-turn dedup, diff-reads and stale-turn summaries only operate past the last cache_control marker (or freely for clients that don't use caching).
- Live cache health monitoring — the dashboard's *Prompt Cache* card shows cache_read vs cache_creation tokens and a Hit Health %. Green (≥80%) means you're paying the minimum possible; red means something is invalidating your prefix.

Full write-up: docs/PROMPT_CACHE.md

How it works

Every request passes through Squeezr on localhost:8080. Compression layers, in order:

1. MCP tool filtering (opt-in) — drop tool definitions from MCP servers you block (mcp_block_servers) or that aren't allow-listed. A single chatty MCP server can cost ~18k tokens…