Launching in 2 weeks. ⏳
Get Early Access Here
Every long session with Claude Code ends the same way: the context window fills up, the agent auto-compacts, and you lose the thread of what it was doing. Tool outputs pile up. grep dumps a wall of text. The conversation gets summarized into mush right when you needed the detail.
DietCode is a drop-in plugin and MCP server that keeps your agent's context lean, automatically, in the background, without you thinking about it, powered by Scaledown.
- 🧠 Intent-aware routing: every prompt gets classified in real time, so the agent reaches for the right tool instead of guessing
- 📉 Automatic context compression: paste in a huge codebase and ask a retrieval question, and the prompt shrinks 50–70% before it ever reaches the model
- 🗜️ Tool-output compaction:
ls,grep,git diff/log/statusand friends get structurally compacted with zero latency; anything still too big goes through Scaledown - ⚡ Progressive proxy compaction: a local proxy that rewrites your outgoing request every turn, keeping a running summary of older turns so your context actually shrinks instead of just growing until it's nuked. This is the one that replaces native auto-compaction entirely.
- 📊 Live savings, right in your status line: watch your token savings tick up in real time (
↓ 125.4K saved · 747 reqs) - 🔧 On-demand tools your agent can call anytime: compress, summarize, classify, and extract, available as first-class tools in every client
- 🔁 Reversible, always: nothing is thrown away for good; pull back the original text behind any summary whenever you need the exact detail
Most compaction is lossy and happens to you, right when it's least convenient. DietCode flips that: compression happens continuously, quietly, before it's ever forced — so your agent stays sharp across long sessions instead of losing the plot every time the window fills up.
We're finishing final testing now. Star this repo to get notified the moment it's live.
Built on Scaledown.