AI assistants
AI assistants
Connect Claude, ChatGPT or another MCP client to your workspace, and build, publish and measure guides by asking for them.
PlayStep has an MCP server, so an AI assistant can work in your workspace with you. Ask it to draft an onboarding guide from your website, fix the step where people drop off, or write this week's report. Anything that would reach your end users waits for a person to approve it in PlayStep.
What you can ask for
| You ask it to | What happens |
|---|---|
| Read guides, analytics, anchors, clips, usage and billing | Runs at once |
| Build and edit guide drafts, add clip links | Runs at once on Build drafts; becomes a request on Suggest changes |
| Generate tours with AI from a public page of your app | Like a draft change, and uses 1 of your monthly AI generations |
| Publish, roll back, pause, resume or unpublish a guide | Always a request that an owner or admin approves |
| Run the workspace: add apps, rotate keys, invite people, set the spend cap | Always a request; only owners and admins see these tools, and only owners the spend cap |
The tool reference lists every tool.
One URL for every assistant
https://api.playstep.app/mcp
It is also on Settings › AI assistants in the dashboard, with a copy button. Add it to your assistant, and the assistant opens PlayStep for you to sign in and choose what it may do. There is no key to paste.
- Connect Claude: Claude on the web, desktop and mobile, and Claude Code.
- Connect ChatGPT: ChatGPT on the web, in Developer mode.
- Other MCP clients: VS Code, Codex, Gemini CLI and the MCP Inspector.
Three layers of control
- Your workspace's rules, in Settings › AI assistants: whether assistants are allowed at all, which ones, the most any of them may do, and who may connect one. Owners and admins set them, and a change applies on every connected assistant's next call.
- The connection. When you connect, you pick a level (Read only, Suggest changes or Build drafts) and the apps it can see. The assistant also works with your role: a member's assistant never sees admin tools.
- Your assistant's own permissions. Claude, ChatGPT and most other clients ask before they use a tool that changes something, and let you allow or block tools one by one.
Permissions and approvals covers each layer.
Changes that reach end users wait for a person
Publishing, rolling back, pausing and unpublishing guides, environment and workspace settings, and deletions always become requests. The assistant gets a review link to give you, and owners and admins approve or reject the request under Settings › Approvals (only owners approve the spend cap). Approving applies the change at once. An assistant can never approve a request, its own or anyone else's.
What an assistant sees
- One workspace, as you. It has your role and sees only the apps you allowed when you connected.
- No secret keys. When an approved request creates or rotates keys, the person who approved it sees the new secret key once, in PlayStep. The assistant gets the publishable keys only.
- Data, not instructions. Guide text, names and page text found by a crawl come from your workspace or the web. We tell the assistant to treat them as data and never follow instructions in them.
Every tool call is in Settings › AI assistants › Recent activity for 90 days: the latest show first, and Show older goes further back.
Next
- Example prompts to copy into your assistant.
- Skills to download, so your assistant follows PlayStep's workflows by itself.
- Build and publish a guide with Claude, a 10-minute tutorial.