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

Sipcode is live on Product Hunt as of today.

What it does

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.

What I measured

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.

Why this matters more than the token count

Anthropic published that cleaner context produces a 29% quality lift and 40% fewer agent errors. The tokens are the proof, reliability is the point.

The dogfood story behind today's launch

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.

Try it

Install:

npm i -g sipcode && sipcode init

PH: https://www.producthunt.com/products/sipcode

Site: https://anuj7411.github.io/sipcode

Repo: https://github.com/Anuj7411/sipcode

Happy to answer questions in the comments.