Plate I · ratio

Why hippo.

— the problem hippo solves, and the things it isn’t.

1 — The problem hippo actually solves

Agentic coding tools forget. Each new session of Claude Code, Aider, or Cursor starts from a blank slate: it has no memory of the architectural decision you made on Tuesday, the name of the function you renamed last week, the failure mode you fixed in the redactor twice last month. You re-explain. The tool re-discovers. You both waste tokens and time.

Hippo is a daemon that watches your shell, your Claude session JSONLs, and your browser (within an allowlist), enriches what it sees with a local LLM, and exposes the result over MCP. Your editor of choice asks ask"why did I rename that function?", "what did I read about probe leakage last month?" — and gets back a synthesized answer with citations. The knowledge stays on your machine. The model runs on your GPU. There is no server.

2 — What hippo is not

3 — Who hippo is for

Developers using agentic coding tools on macOS arm64. That’s the audience: it’s narrow on purpose. If you don’t use Claude Code (or a similar tool), you have very little to gain from running hippo today. The shell-capture path works for any zsh user, but the synthesized value — cross-session memory for an agentic tool — comes from feeding the tool back into itself.

4 — What hippo will never do

If you spot any drift between the bullets above and the implementation, please open an issue. Concordantia is load-bearing.