Use Setgreet with AI agents

There are three ways to build with AI on Setgreet:

• In-product AI -- describe a flow in the dashboard and the assistant drafts and edits native screens for you. No setup. See Build with AI.
• MCP server -- connect Claude, Cursor, Codex, or any MCP client to your workspace and let it author, publish, and analyze flows through your existing tools. That is this page.
• Agent instructions file -- a drop-in block that teaches your coding agent to install the SDK and use the MCP server correctly in your own repo. See Setgreet for AI agents.

The MCP server lets AI agents read and write your Setgreet workspace -- flows, screens, analytics, experiments, segments, theme, localization, billing, team, and integrations -- through a single OAuth connection. It implements the Model Context Protocol, the open standard supported by Claude, Cursor, VS Code, Windsurf, Codex, and a growing list of other clients. Setgreet hosts the server at https://mcp.setgreet.com/mcp -- there is no install, no local process, no API key to manage.

Install

Pick your client. Each tab is a copy-paste install. The hosted endpoint is the same in every case; only the client config differs.

{
  "mcpServers": {
    "setgreet": {
      "url": "https://mcp.setgreet.com/mcp"
    }
  }
}

claude mcp add setgreet --transport http --scope user https://mcp.setgreet.com/mcp

{
  "mcpServers": {
    "setgreet": {
      "url": "https://mcp.setgreet.com/mcp"
    }
  }
}

{
  "servers": {
    "setgreet": {
      "type": "http",
      "url": "https://mcp.setgreet.com/mcp"
    }
  }
}

code --add-mcp '{"name":"setgreet","type":"http","url":"https://mcp.setgreet.com/mcp"}'

{
  "mcpServers": {
    "setgreet": {
      "serverUrl": "https://mcp.setgreet.com/mcp"
    }
  }
}

[mcp_servers.setgreet]
url = "https://mcp.setgreet.com/mcp"

https://mcp.setgreet.com/mcp

Authorize

The first time your agent calls a Setgreet tool, the MCP server opens a browser to the Setgreet dashboard. You sign in with your existing account, the browser hands the redirect back to the client, and the agent prints your user, organization, role, and granted scopes.

Setgreet uses OAuth 2.1 with PKCE -- there is no shared secret to leak, and the agent never touches your refresh token. Access tokens are short-lived (5 minutes) and refresh transparently in the background; refresh tokens rotate on every use and stay valid for 30 days.

You can list and revoke active MCP sessions from the dashboard at Settings -> MCP Connections (/settings/mcp-connections). Revoking a session takes effect immediately -- the agent's next tool call gets authentication_required and prompts re-authorization.

To verify the install, ask the agent:

Call the setgreet whoami tool.

It returns your user, org, role, granted scopes, and token expiry.

How scopes work

Scopes are not requested by the client. They are derived from your dashboard role at token issuance and written directly into the JWT. A viewer's token literally cannot carry flows:write -- the scope is absent from the scope claim, so even a malicious client cannot use it. Admins can create segments and manage team; viewers cannot. Role demotions take effect on the next token refresh (within 5 minutes).

Destructive operations (delete, unpublish, complete experiment, remove member, regenerate share token) require an explicit confirm: true parameter -- the agent has to ask you before passing it. If it doesn't, the MCP server refuses the call with confirmation_required before the API is touched.

End-user PII (email, phone, name, address, IP) is partitioned behind a separate endusers:read:pii scope granted to admins and above. Without it, the backend silently replaces those fields with "[redacted]" -- enforced server-side, not at the MCP layer.

Plan availability

MCP access is included on every plan, including Free. See pricing.

What's next

• Recipes -- end-to-end agent workflows with copy-paste prompts.
• Setgreet for AI agents -- the drop-in instructions file for coding agents.
• Tools -- the full tool catalog grouped by domain.
• Troubleshooting -- error codes (authentication_required, scope_required, plan_required, confirmation_required) and how to resolve each.
• llms.txt and llms-full.txt -- machine-readable maps of these docs for agents.