Skip to content

The engine for autonomous software production

zeroshot is an open-source workflow engine for coding agents. Write your team's development and QA process as a graph of implementation, independent review, testing, repair and delivery. Agents do the work; the graph decides what happens next, and the engine enforces its transitions and retry limits.

goalworkeracceptancereviewcode reviewboth mustapprovePRCImergepassnofailrepairagainyou, reading the diffnot anymoreadversarialtestissue #4225writesjudgesbuilt inboots the stackIndependent review goalissue #4225workeracceptancereviewcodereviewboth mustapprovePRCImergepassfailnorepairagainyou, readingthe diffadversarialtestwritesjudgesbuilt inboots the stackIndependent review
An example graph: independent reviews and an adversarial test before the PR, CI before merge, and every rejection back through repair. More examples on zeroshot.sh

Start here

  1. Install the CLI:

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

  2. Run a first task with the built-in software-change graph.

  3. Build a review loop: write your own graph and runtime plan.

Working in ChatGPT, Codex or Claude? The plugin runs zeroshot without the CLI. Prefer a browser? zeroshot ui edits profiles, follows live runs and replays finished ones.

What a run needs

  • A graph: the steps, the routes between them and the typed state they share.
  • A runtime plan: the agent CLI, provider and model for each step, and the connections it needs.
  • An input: the task. zeroshot checks it against the graph before any agent starts.

Model output cannot add an edge or raise a bound. zeroshot records every step's result as the run advances, and closing the terminal does not stop the run. How execution works

Built-in graphs

Run them as they are, or export one as the start of your own.

Graph What it does
software-change Implements the task, runs acceptance and code review in parallel, repairs what they reject, then delivers.
single-worker One agent, no review loop.
pr-review Reviews a pull request.
code-review Reviews a pull request in depth.
auto-research Runs ten rounds of experiments by default. A new result replaces the current one only when the progress judge accepts it.

Where it runs

Target Where code and run state live
Local Any folder on your machine (--cwd FOLDER). Git is optional without delivery. Install
Self-hosted target Infrastructure you control. The target image checks out the source; agents and the run ledger stay there. Targets
zeroshot Platform A managed target that takes the same graph and runtime plan, with accounts, organization policy and queues. Platform docs

Where code is processed

The target and the model provider are separate choices. Wherever a run happens, the agent sends its model provider the code and context it needs, so self-hosting zeroshot does not make model processing local. Check the provider's data-handling terms before running on client code. Runtimes and connections covers providers and credentials; for Platform, see the privacy notice and retention.

Reference

  • CLI, generated from the command tree
  • Python API, generated from the SDK
  • Cluster API, generated from the OpenRPC contract

These pages follow main and may describe unreleased changes. For an installed release, pick its version in the header. Documentation versions