Your repo stays in your copilot
Your assistant (Claude Code, Copilot, Codex, Cursor) holds the repository and decides what a request needs. Pauhu never copies your repo.
How it works
Repo in your copilot. Working code back.
MCP is at once a local client and a hosted server, so your code stays on your machine while you use a hosted compiler. Only the part Pauhu needs crosses the wire; your repository stays where it is.
Only the part it needs crosses
Your copilot sends Pauhu only the small piece of context the request needs. Not the whole repo, not your secrets.
The hosted compiler builds the code, no model tokens
Pauhu builds the code from published standards and documented APIs. No model call, no GPU run on the answer it returns.
Code back, with a receipt
You get working code in your copilot, and a receipt logs the tokens and energy the call avoided. Your model handles the intent.
The same code, every time
A compiler your copilot can call.
Your copilot was trained on the whole inconsistent public corpus, so it predicts code token by token and sometimes guesses wrong. Pauhu hands it the tool it lacked: the same request returns the same code every time. Composed, not guessed, because it is a finite-state interpreter of grammar and published, cited operations, not a probabilistic model.
The library keeps growing. Anything Pauhu does not cover, your own assistant handles as it does today, and the receipt shows which calls Pauhu built at no token cost.
Your copilot, holding the repo
Just the part it needs, not the repo
mcp.pauhu.dev builds it from published sources
Working code, 0 model tokens
Receipt: tokens and energy saved