The agent that writes the code should not be the one that decides it works.

Independent review and repair for coding agents.

npm install -g @the-open-engine-company/zeroshot

GitHub Docs

running Illustrative run software-change

task “Add JSON output to the status command and cover it with focused tests.”

00:00runstarted · software-change
00:01workerimplementing the task
00:46workerdone · 3 files changed
00:46reviewacceptance + code, in parallel
01:14codeapproved
01:21acceptancerejected · empty-list case has no test
01:22repairfixing with feedback from both reviews
01:50reviewacceptance + code, round 2
02:17acceptanceapproved
02:20codeapproved
02:21delivery--pr · pull request prepared, not merged
02:33runpassed · configured checks accepted

We built Zeroshot because we were tired of being gaslit by agents telling us broken code was ready.

One agent, grading itself

agentwrites the code
agentdecides what “done” means
agentchecks its own work
agent“It’s ready.”

Zeroshot’s built-in graph

workerwrites the code
acceptancechecks it against the task
codereviews the change, independently and in parallel
repairfixes rejected work with both reviews’ feedback, then both review again
deliverywaits until the graph’s checks pass

The implementing agent never approves its own work. Zeroshot doesn’t replace Claude Code, Codex or GitHub Copilot. It runs one of them as the worker and as the reviewers.

A passing run means its configured checks accepted the work. How much that covers depends on the requirements, reviewers, tests and environment you provide. Watch a single agent vs. Zeroshot (scripted video).

Run your first task

Install

Needs Node.js 18 or newer. Installs a verified native binary for Linux and macOS (x64, arm64) or Windows x64.

You also need Codex, Claude Code or GitHub Copilot installed and signed in. Local runs can reuse that login, including subscription-backed sessions.

npm install -g @the-open-engine-company/zeroshot
Describe the task

Put the change you want in input.json, and the harness and model in runtime.json. This example uses Codex: replace YOUR_MODEL_ID with a model your Codex CLI supports.

Other harnesses and providers
input.json
```
{
  "task": "Add JSON output to the status command and cover it with focused tests."
}
```
runtime.json
```
{
  "harness": "codex",
  "provider": "openai",
  "model": "YOUR_MODEL_ID",
  "effort": "high"
}
```
Run it

The worker edits the current Git worktree, so start in a clean one meant for the task. Without a delivery flag, the work stays local. Add --validate-only to check the graph, runtime and input first.

To watch it, run zeroshot ui in another terminal, open http://127.0.0.1:4173/ui/ and go to Runs. Or use zeroshot status RUN_ID and zeroshot logs RUN_ID.
shell
```
zeroshot run \
  --title "Add JSON status output" \
  --template software-change \
  --input input.json \
  --uniform-runtime-config runtime.json
```

An explicit graph, not a longer prompt

The graph defines the sequence, the parallel steps, the retry paths and the exit conditions before the run starts. Runs are bounded, and events are recorded in a durable SQLite ledger.

zeroshot template show: Start from the built-in graph and add your own stages, like adversarial tests or E2E.; Graph contract
--runtime-config: Give each node its own model and effort instead of one for the whole run.; RuntimePlan
--profile: Save a graph and its runtime settings as a profile. The CLI and browser UI share them.; UI setup
--push --pr --ship: Ship as far as you trust it: a pushed branch, a mergeable PR, or PR, CI and merge.; CLI reference
zeroshot target add: Run graphs on a self-hosted Docker target, on infrastructure you control.; Target image
auto-research: Bounded rounds of experiments, judged independently on evidence, method and progress.; Execution

Try it on your next task.

Starred by engineers at Google, Meta, Shopify, Uber, Booking.com, GitHub, Cloudflare, Atlassian and JetBrains. Questions, ideas, or a run worth showing? Join the Zeroshot community on Discord.