AgentKrew

Claude Code runtime

Claude Code is AgentKrew's default runtime. The Claude adapter renders canonical artifacts into Claude-native project files — no settings or CLAUDE.md edits required.

Install

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

(--runtime claude is the default, so a bare npx @agentkrew/cli init does the same.) Verified against Claude Code 2.1.128.

What init generates

text
.agentkrew.json                       installation manifest
AGENTKREW.md                          project context (create-only)
.claude/agents/<name>.md              one subagent per kit agent
.claude/commands/<name>.md            one slash command per kit command
.claude/skills/<name>/SKILL.md        one directory per kit skill
.claude/skills/workflow-<name>/SKILL.md  one per kit workflow

For the Engineer Kit that is 15 agents, 7 commands, 18 skills, and 3 workflow- skills (45 files total).

Rendered agents carry Claude frontmatter (name, description); the prompt body is the kit's prompt.md plus a pointer to AGENTKREW.md. Workflows render as commands — with the workflow- prefix so a command and a workflow of the same name (for example review) never collide.

How you use it

Commands are slash commands. Open Claude Code in your project and:

text
/create-feature "Add a user profile page"
/fix-bug "CSV import fails on empty rows"
/review
/ship

Workflows are slash commands too, via their generated skill:

text
/workflow-feature-development "Add a user profile page"

Agents are subagents. Claude Code reads each agent's description and delegates matching work to it automatically (or name it explicitly with @tech-lead). Workflow plans embed "spawn one agent per step" instructions, including parallel groups, so the whole team runs inside the session.

Skills load on demand. Kit skills are picked up by description match when an agent needs them; the skill directory name is also a slash command.

Project context

agentkrew init creates AGENTKREW.md at the project root and never overwrites it. Every generated agent carries a pointer to it, so all agents read one shared context file instead of duplicating it.

Repo-root CLAUDE.md is user-owned: AgentKrew never writes it. If you keep project instructions there too, keep them short and let AGENTKREW.md hold the AgentKrew context.

What init never touches

  • Repo-root CLAUDE.md
  • Anything outside .claude/, .agentkrew.json, and AGENTKREW.md

agentkrew update refreshes only generated files under .claude/; your edits to AGENTKREW.md survive every update. See customization and updates.

Verify

bash
npx @agentkrew/cli doctor
text
✓ Claude detected (2.1.128)
✓ Runtime version compatible
✓ AgentKrew manifest valid
✓ 15 agents
...
AgentKrew is healthy.

Notes and limits

  • Name shadowing: Claude Code lets a skill and a command file claim the same /<name> (the skill wins). agentkrew validate flags any skill/command overlap so the shadowing is never silent.
  • No runtime APIs involved: everything is plain files, so a fresh session discovers the kit with no auth, model calls, or settings changes.
  • Design notes and version-specific research: claude-research.md.