Skip to content

Core documentation

Build and review software through a graph

A Zeroshot graph records implementation, review, repair, and delivery as bounded execution steps. Its authored transitions determine which node runs next.

Run a first task Read the CLI reference

Open the workspace UI to design profiles, watch live runs, and replay recorded history.

What a run contains

Three authored values determine execution:

  • a graph that defines control flow and typed state;
  • a runtime plan that binds each executable node to a harness, provider, model, and named connections;
  • an initial input that belongs to the caller and that Zeroshot checks against the graph before execution.

After validating those values, Zeroshot opens a durable ledger where it records each node result. Authored edges then route review, retries, and the final result.

flowchart LR
  I[Goal and source] --> A[Admission]
  A --> W[Worker]
  W --> R1[Acceptance review]
  W --> R2[Code review]
  R1 --> D{Both accept?}
  R2 --> D
  D -- no --> F[Bounded repair]
  F --> R1
  F --> R2
  D -- yes --> X[Result or delivery]

The built-in software-change template uses the route shown above. For work that does not need its review loop, use single-worker. For ten rounds of measured, reversible research against a durable charter, use auto-research; custom graphs follow the same protocol contracts.

Choose where it runs

  • Local

Local mode works in the current Git worktree, keeps state on the same machine, and calls your chosen harness.

Install Zeroshot

  • Self-hosted target

The target image checks out source on infrastructure you control. Agent processes and the run ledger stay there as well.

Understand targets

  • Zeroshot Cloud

Cloud accepts the same graph and runtime plan through a managed target. Its documentation covers accounts, organization policy, and queue behavior.

Open the Cloud docs

Read more

How execution works describes the graph and reduction model. Read Runtimes and connections before choosing models or supplying credentials; Observe and control runs covers failed-run recovery, durable status, and log streams. Use a local graph as an ACP agent covers the experimental stdio endpoint. Build a review loop builds a custom graph and runtime plan step by step.

These reference pages are generated from product-owned definitions:

  • CLI reference, generated from the Clap command tree;
  • Python API, generated from the SDK's public objects and docstrings;
  • Cluster API, generated from the checked-in OpenRPC contract and protocol schemas.

Current follows main and may describe unreleased changes. For an installed release, choose its minor version from the header; each minor documents its newest published patch. Documentation versions defines the URL and manifest contract.