Build AI workflows with AgentKit
Install AgentKit, add a kit to your coding assistant, and run your first workflow.
AgentKit brings reusable AI workflows to coding assistants such as Claude Code
and Codex. The ak CLI installs each kit for the runtime you use, preserves
recovery points around managed changes, and keeps release verification in the
install and update path.
Choose your starting point
- New to AgentKit? Start with Installation, then complete Onboarding.
- Already signed in? Follow the Quickstart to install Engineer Kit and run one skill.
- Looking up exact syntax? Open the CLI reference, generated from the released command tree.
Install AgentKit
Clone the repository and verify your environment:
# Clone repository
git clone https://github.com/saccerwin/agent-kit.git ~/.agentkit
cd ~/.agentkit
# Run environment & runtime diagnostics
node bin/ak.js doctor
# Optional: Add `ak` to your PATH
ln -sf ~/.agentkit/bin/ak.js ~/.local/bin/akOn Windows PowerShell:
git clone https://github.com/saccerwin/agent-kit.git $HOME/.agentkit
cd $HOME/.agentkit
node bin/ak.js doctorContinue with Installation for full runtime projection details and options.
Run your first workflow
After you authenticate and install Engineer Kit, invoke the same skill with the syntax required by your assistant:
| Claude Code | Codex |
|---|---|
/ak:cook Add a health-check endpoint | $ak:cook Add a health-check endpoint |
See Quickstart for the complete sequence and expected checks.
Choose a kit
- Engineer Kit is for software work: clarify requirements, plan changes, implement, test, review, and troubleshoot.
- Marketing Kit is for marketing and content workflows.
Choose a kit by the work you want to do. You invoke individual skills only after the kit is installed for your runtime.
Recover or get unstuck
AgentKit previews managed installation changes and takes recovery snapshots
before supported destructive writes. Keep any backup ID printed by ak, and do
not use --force until you understand the ownership conflict. Start with
Onboarding recovery
or use the CLI reference to inspect the exact command.