Builder · Public · built in 2026 · I run it on tunyt daily
Zerd.si
I built zerd to make coding agents show their work before they say a task is done. It runs project checks, asks a second model to review the diff, verifies findings and checks every cited file path on disk. It is a native Rust binary, and I use it on tunyt. In one head-to-head on 11 small tasks, it tied Claude Code at 10/11—not a win, but a result I can show.
- Cached input in a real run
- 85%one 21-call session on tunyt's web app
- Against Claude Code
- 10/11 each11 small tasks, one run each, different models. Same range, not a win.
How it works
- Your checks run on the changed files. A failure goes back to the model as a repair request.
- A reviewer on a different model reads the real diff, and a verifier tries to disprove each finding. A confirmed HIGH gets run in a throwaway copy of the change.
- Every path and line number in the answer is checked on disk before you see it.
Why it's cheap
Agents re-send the whole conversation on every call. zerd tracks which lines of each file the model has already seen and only sends what's new, and it finds code with ranked search and outlines before reading anything. It runs Claude, GPT, Codex, DeepSeek and anything behind an OpenAI-style API, and routes each request to a cheap or strong model by difficulty.
Safety
The rules live in code. File tools stay inside your folder, shell commands get scanned before they run, and on macOS an OS sandbox sits under both. zerd snapshots the working tree before each request, so /undo works.
What it can't do yet
No image input and no IDE extension, and the OS sandbox is macOS only. On Linux and Windows you get the confined file tools and the shell path scan. My one head-to-head against Claude Code on 11 small tasks ended 10/11 each, so I don't claim it wins. I use it every day on tunyt. That's my evidence.