AgentKrew

Getting started

Install AgentKrew into an existing project, verify it, and run your first multi-agent workflow. The whole flow takes a few minutes.

Prerequisites

  • Node.js >= 22 (for npx; no global install is required).
  • One supported coding runtime on your machine:
  • A project (any language) with a Git repository.

1. Install

Run the initializer from your project root:

bash
npx @agentkrew/cli init

By default this runs interactively. To skip the prompts:

bash
npx @agentkrew/cli init --runtime claude --kit engineer --yes

Init options

Option Values Default
--runtime <runtime> claude, codex, opencode claude
--kit <kit> engineer, marketing, cross-kit required
--kit-source <source> license, local license
--kit-path <path> local kits directory (required for local) —
--yes skip confirmation prompts off

Kit sources

  • license (default) — downloads the kit from the AgentKrew registry. Requires the AGENTKREW_LICENSE_KEY environment variable. The registry URL defaults to https://api.agentkrew.dev and can be overridden with AGENTKREW_REGISTRY_URL.

  • local — installs from a local kits/ directory; intended for contributors and for evaluating AgentKrew from a source checkout. --kit-path is required:

    bash
    npx @agentkrew/cli init \
  --kit engineer \
  --kit-source local \
  --kit-path ./kits

2. What init writes

Given --runtime claude --kit engineer, init shows you the full file list before writing anything, then creates:

text
.agentkrew.json                    installation manifest (who/what/when)
AGENTKREW.md                       project context — yours to edit
.claude/agents/<name>.md           one file per agent        (managed)
.claude/commands/<name>.md         one file per command      (managed)
.claude/skills/<name>/SKILL.md     one file per skill        (managed)

The directory name follows the runtime: .claude/, .codex/, or .opencode/.

Files fall into three ownership classes — managed, user-owned, and generated — and only managed files are ever refreshed by agentkrew update. See customization for what you can safely edit.

Describe your project

Init finishes with:

text
Created AGENTKREW.md
Please describe your project and conventions.

Open AGENTKREW.md and fill in your project overview, tech stack, and conventions. Every agent reads it, so this one file is the highest leverage customization you can make. Updates never overwrite your edits to it.

3. Verify the installation

bash
npx @agentkrew/cli doctor

A healthy project prints:

text
AgentKrew Doctor

Project root: /path/to/your/project

✓ Claude detected (2.1.128)
✓ Runtime version compatible
✓ AgentKrew manifest valid
✓ Engineer kit installed
✓ No broken references
✓ Runtime configuration valid
✓ 15 agents
✓ 18 skills
✓ 7 commands
✓ 3 workflows
✓ AGENTKREW.md found

AgentKrew is healthy.

doctor exits non-zero when anything is broken, so it is safe to use in scripts and CI. See troubleshooting when it reports a failure.

4. Run your first workflow

Open your project in your runtime and invoke an AgentKrew command. With Claude Code the installed commands are slash commands:

text
/create-feature Add a user profile page

That single command drives the feature-development workflow: a tech-lead plans the work, specialists implement it in parallel, then test-automator, security-auditor, code-reviewer, and shipper verify it. Other starting points:

text
/fix-bug        reproduce, root-cause, fix, regression test, review
/review         aggregate code, security and test review findings
/ship           gate the release on real checks

Command invocation differs slightly per runtime — see OpenCode, Codex, or Claude Code.

5. Explore what is installed

bash
npx @agentkrew/cli list agents --installed       # what teams you have
npx @agentkrew/cli list skills --installed
npx @agentkrew/cli list commands --installed
npx @agentkrew/cli list workflows --installed
npx @agentkrew/cli list agents --installed --json # the same rows as JSON
npx @agentkrew/cli info tech-lead --installed    # one artifact in detail
npx @agentkrew/cli info tech-lead --installed --json  # machine-readable

--installed reads your project's installation. Without it, list and info read the source kits that ship with a source checkout of AgentKrew (contributor workflow). --json works on list itself or on any subcommand (list --json agents or list agents --json), and every row carries the kit the artifact came from.

Next steps