because the agent that wrote the code shouldn't be the one that says it works.
A meta-harness that enforces review loops on your coding agent. Like CI, except the checks are agents and a red result goes back for repair.
Getting started? Ask your agent to pick a small issue in your repo and ship it with Zeroshot.
An npm CLI. Free and open source. · Docs →Star on GitHub ↗Docs ↗
- Meta
- Shopify
- Uber
- Booking.com
- GitHub
- Cloudflare
- Atlassian
- JetBrains
-
Because we were tired of being gaslit by agents telling us broken code was ready.
-
No. The reviewers and tests you set up do, and a run only passes when they accept it. Want to look anyway? Use
--pr. -
Keep using them. Zeroshot runs your agent as the worker and as separate reviewers, so the one that wrote the code never approves it.
-
Sure. A skill is advice, so the agent still decides whether to believe the reviewer it spawned and when it’s done. A Zeroshot graph is fixed before the run starts, so every run goes through the same reviews and ends when they accept or the retries run out.
-
No. Locally it runs on the Codex or Claude Code login you already have, subscription included. You can also use keys for OpenAI, Anthropic, OpenRouter or Amazon Bedrock, with a different model per agent. docs ↗
bring your own graph topology.
Your repo, your rules. Add reviewers, tests and repair loops, or rewire the whole graph.
what a design slop reviewer catches.
The experiment: Claude Code on Opus 5.5 builds Zeroshot’s landing page from its README, on its own and inside a custom graph with a design slop reviewer. To keep it a design test, that graph skips code review.
one session, no reviewer
plain Claude Code
Its final report
“I checked it with headless-Chrome screenshots at phone, tablet and desktop widths, in light and dark mode, and with reduced motion.”
worker → design slop reviewer → repair
a custom graph for this experiment
Round 1: sent back
- Too long: the README poured into ten screens.
- It explains how Zeroshot works five times.
- Every section looks the same, so nothing stands out.
- Apart from the demo, it could be any dev tool’s page.
- The install command, the one line people copy, breaks in two.
- On phones the staged demo run loses its “illustrative” label and passes as real.
Fix: make the rejection the page, with the run up top and the rest cut by half.
Round 2: accepted
Five screens instead of ten.
zeroshot cloud.
Put your topology in the cloud and enforce it on every run your team starts from an issue.
just you
Start a run from a GitHub issue, even from your phone, and close the laptop. Nothing merges until your reviewers say yes.
your team
Your standards live in your topology, not in anyone’s skills folder or anyone’s head. Every issue run goes through it.
Comment @zeroshot-cloud run on a GitHub issue.Zeroshot Cloud ↗Cloud docs ↗
Free for one run at a time, then $100 a month for each extra run in parallel. Models run on your own API keys.
what early adopters say.
“You have really allowed me to just… create more things in my personal time than I have before.”
tester running Zeroshot on Claude and Codex
“It’s really nice to be able to customize the nodes and build your own loop.”
tester, Kubernetes platform, translated from Norwegian
“you guys are solving this in a way that is more satisfying than the google ADK (to me)…”
Discord community member
not just
for coding.
Autoresearch runs experiments through three independent judges.
Explore autoresearch ↗
Watch a research run · 19 sec
stay in the loop.
Star it on GitHubJoin the Discord
design partners get an exclusive hoodie :)
who's building it.
Three co-founders. Eivind and Michael met doing machine learning PhD research at the Technical University of Munich. Olle has 7 years in startup growth and runs our brand.