GitHub - the-open-engine/zeroshot: Your autonomous engineering team in a CLI. The agent loop produces senior-level code that you can actually trust in prod because of non-negotiable feedback from independent reviewers. Supports Claude Code, OpenAI Codex, OpenCode, and Gemini CLI with trivial setup.

The agent that wrote the code shouldn't be the one that says it works.

Zeroshot is an open-source, multi-agent orchestration engine for autonomous software engineering. It drives a coding agent you already run (Claude Code, OpenAI Codex, Gemini CLI, or OpenCode) through an executor-verifier loop: an agent writes the change, then an independent verifier that never saw how it was made approves it, or hands back a reproducible failure. The loop runs until the change is verified.

Install

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

Requires Node ≥ 18 and at least one provider CLI (Claude Code, Codex, Gemini, or OpenCode). Linux and macOS today; Windows is deferred.

One run, unattended. 100× speed · 90-minute run · 5 iterations to approval.

How it works

Zeroshot separates the agent that writes the code from the agent that judges it.

A conductor sizes the workflow to the task. An executor (an AI coding agent) implements the change in an isolated workspace (git worktree or Docker). Then an independent verifier inspects the result without ever seeing the executor's context or history, so it cannot approve its own reasoning. The verifier returns APPROVED, or REJECTED with an actionable, reproducible failure, and the loop repeats until the change is verified or hands back a concrete reason it isn't. Every step is written to a crash-safe SQLite ledger, so a run survives a reboot and resumes where it stopped.

task --> plan --> implement --> verify --> APPROVED --> done
                      ^            |
                      +- REJECTED -+   (reproducible failure)

Bring your own provider and your own backend. Zeroshot orchestrates the agents that write your code; it doesn't store your keys or replace your models.

How is this different from a single coding agent?

Quick start

zeroshot run 123                 # a GitHub issue number
zeroshot run feature.md          # a markdown spec
zeroshot run "Add a --json flag" # inline text

Describe a non-trivial task inline and let the loop run it to a verified change:

zeroshot run "Add optimistic locking with automatic retry: when updating a user,
retry with exponential backoff up to 3 times, merge non-conflicting field changes,
and surface conflicts with details. Handle the ABA problem where version goes A->B->A."

# Run
zeroshot run <input>            # issue number / URL / key / markdown file / inline text
zeroshot run 123 --worktree     # isolate in a git worktree
zeroshot run 123 --docker       # isolate in a container
zeroshot run 123 --pr           # worktree + open a pull request
zeroshot run 123 --ship         # worktree + PR + auto-merge on approval
zeroshot run 123 -d             # background (daemon)
zeroshot run 123 --provider gemini   # override the provider for this run

# Monitor & manage
zeroshot list                   # all clusters (--json)
zeroshot status <id>            # cluster details
zeroshot logs <id> -f           # stream logs
zeroshot resume <id> [prompt]   # resume a stopped/failed run
zeroshot stop <id>              # graceful stop
zeroshot kill <id>              # force kill
zeroshot export <id>            # export the conversation

# Library & config
zeroshot providers              # list providers / set-default / setup
zeroshot agents list            # available agents (agents show <name>)
zeroshot settings               # view / get / set settings
zeroshot cmdproof check <id>    # reuse a verified command result

Providers and backends

Zeroshot shells out to provider CLIs; it stores no API keys and manages no auth. Pick a default and override per run.

zeroshot providers                    # see what's installed
zeroshot providers set-default codex
zeroshot run 123 --provider gemini

Issue backends are auto-detected from your git remote: GitHub, GitLab, Jira, and Azure DevOps. Paste a number, key, or URL:

zeroshot run 123                                              # GitHub
zeroshot run https://gitlab.com/org/repo/-/issues/456        # GitLab
zeroshot run PROJ-789                                         # Jira
zeroshot run https://dev.azure.com/org/project/_workitems/edit/999  # Azure DevOps

Each backend needs its own CLI installed (gh, glab, jira, or az). See docs/providers.md for model levels and setup.

Isolation

By default, agents modify files only; they do not commit or push. Opt into isolation to let the loop own a branch (the flags cascade: --ship → --pr → --worktree).

zeroshot settings set dockerMounts '["gh","git","ssh","aws"]'
zeroshot run 123 --docker --mount ~/.aws:/root/.aws:ro
zeroshot run 123 --docker --no-mounts

Scope and status

Zeroshot performs best when a task has clear acceptance criteria. If you can't say what "done" means, an independent verifier can't confirm it.

• Pre-1.0 in spirit. Interfaces still move between releases; pin your version. (The npm version auto-increments on every merge, so read it as a build counter, not a stability promise.)
• Crash-safe. All state persists to a SQLite ledger; zeroshot resume <id> continues at any time.
• No TUI in this release. Monitor with zeroshot logs <id> -f, zeroshot list, and zeroshot status <id>.

Zeroshot is Layer 01 · Verification of The Open Engine, the open stack for autonomous software production. Generating code is easy; trusting it is not. The engine is layered because trust is layered:

Zeroshot runs the loop: an agent writes the change, and an independent verifier decides whether it holds: approve, or a reproducible failure. Opcore is the sibling layer, a deterministic, local, read-only constraints gate for coding agents (currently private alpha 0.1.0-alpha.0, built in the open, not yet published). Verification asks "does this meet the goal?"; constraints ask "is this within tolerance?"

Each layer ships the same way: extracted from the platform we run, then opened. Trust nothing. Verify everything.

Contributing

See CONTRIBUTING.md for development setup, CODE_OF_CONDUCT.md before participating, and SECURITY.md for security reports. More in docs/ and CLAUDE.md.

Questions and help: Discord.

License

MIT. The Open Engine Company.