Get Started
Agent Skill
Claude Code and other coding agents that read skills can add the agent workflow to your app for you. The skill tells the agent to read your app before it changes anything, install from our registry, wire the components on one event log, and leave the rest of your code alone.
Install the skill
Run this from your project's root. It puts the skill where Claude Code looks for project skills:
mkdir -p .claude/skills/agents-and-humans .claude/skills/agents-and-humans/references for f in SKILL.md references/boundaries.md references/contract.md references/inspect.md references/install.md references/wiring.md; do curl -fsSL https://agentsandhumans.ai/skills/agents-and-humans/$f -o .claude/skills/agents-and-humans/$f done
Then ask your agent:
Use the agents-and-humans skill to add the agent workflow to this app.
The files are plain Markdown: the file list,
SKILL.md, references/boundaries.md, references/contract.md, references/inspect.md, references/install.md, references/wiring.md.
The source is in the repository, under the MIT licence.
The five steps
- Reads your app first. The framework, where shadcn puts files, your design tokens and dark mode, the components you already have, and how your server knows who is signed in. It shows you a short note and the list of files it plans to add before it writes anything.
- Installs from the registry. One command brings the bundle and the components it needs.
- Wires them on the contract. One run id, the plan id on the approval, the plan step id on each trace step, and one log your server keeps.
- Keeps the boundaries. Your server verifies each approval against the request it issued, and redacts secrets from each trace frame before it leaves the server. The agent changes only the files on its list.
- Proves it. It runs your typecheck and build, walks the example page, and reports the files it touched and what is left for you.
The bundle it installs
Plan Preview, Approval Gate, Progress Trace and a Review Receipt on one Interaction Contract log, with a working example page and its server route. The person accepts the plan, approves the request your server issued for it, watches the run and reviews the receipt. Your server verifies each approval and redacts secrets from the trace.
npx shadcn@latest add https://agentsandhumans.ai/api/registry/agent-workflow.json
After the install, run your app and open /agent-workflow. The example page and its server
route run as they are; the server keeps its approvals in memory and names a demo person, so replace both
with your database and your sign-in before real use. The
Storybook page shows the same flow.