I shipped Sipcode today: keeping Claude Code's context clean for sharper answers

Sipcode is live on Product Hunt as of today. Local proxy plus MCP server for Claude Code. It rewrites tool output (Read, Bash, Grep) before the model sees it, removing redundancy without dropping information. On my own 3,567,170-token dogfood corpus: 62.6% median tool-output savings, range 37.4 to 80.6%, $67.43 saved. 0 network calls. 1,363 tests. 15 MCP tools. MIT. Anthropic published that cleaner context produces a 29% quality lift and 40% fewer agent errors. The tokens are the proof, reliability is the point. Last week I caught my own drift tool reporting 624,940 tokens wasted in a 4-hour session. My proxy --stats credited only 7,553 saved on the same session. 83x undercount, by my own tool. Root cause was mid-session installs leaving the dedup cache empty for the first half of the session. Shipped v1.6.15 with Verified Warm-Fill 24 hours later. Drift now reads "no drift detected." Shipped v1.6.16 today with the cache-defer and grep-cap fixes. Three releases in nine days, each closing a real bug I caught dogfooding. Install: npm i -g sipcode && sipcode init PH: https://www.producthunt.com/products/sipcode https://anuj7411.github.io/sipcode https://github.com/Anuj7411/sipcode Happy to answer questions in the comments.