Trust

Security

Last updated: October 31, 2026

MCP Tutor is a connector your AI assistant (Claude, ChatGPT, Gemini, Grok or Perplexity) calls to fetch training content and record your progress. This page explains how that connection is secured and what data it can reach. For what we collect and who processes it, see the Privacy Policy.

How an assistant signs in

  • OAuth 2.1 with PKCE. Every assistant app connects through the standard authorization-code flow. PKCE is required and only the SHA-256 (S256) method is accepted. You approve the connection on an MCP Tutor page that shows exactly where the authorization will be sent. The one exception is the optional command-line token described below.
  • Exact callback addresses. A sign-in code is only ever delivered to an address registered for that app, one of Claude’s own hosted callbacks, or a loopback address on your own machine (which command-line tools use). Codes are single-use and tied to the address that requested them, and to the app when the request names one.
  • Short-lived tokens. The access tokens your assistants receive expire after one hour. Refresh tokens last up to 30 days and are replaced every time they are used, so a copied refresh token stops working after its next use.
  • One long-lived exception. Command-line tools that can only send a fixed header (Gemini CLI and Antigravity CLI) use a permanent connector token instead of signing in. It works only on the Gemini address, does not expire, and is shown on your dashboard once you are signed in. Treat it like a password, and email support@mcptutor.com if it is exposed.
  • Disconnect any time. “Disconnect all connectors” on your account page signs every assistant out by revoking the tokens issued to them. It does not change the command-line token above. Tokens are only accepted in the Authorization header, never in a URL.

What the connector can do

  • Read curriculum, playbook and knowledge content, and read and record your own progress, demonstration attempts and playbook runs.
  • Every tool is labelled read-only or write, and tools that change shared data are marked as such, so your assistant can ask before acting.
  • Managers and admins can read their own team’s progress. Members of one organization can never read another organization’s data.
  • MCP Tutor never reads your chat history or your files. It receives only what your assistant passes to its tools — for example, your demonstration response and the assistant’s assessment of it.

How data is kept separate

  • Every account belongs to one organization. The server scopes reads and writes to that organization in application code, and row-level security in the database adds a second layer for direct access from the browser.
  • MCP Tutor staff can build training for a client organization only through audited tools, and see that organization’s progress only in aggregate — never an individual learner’s identity or answers. Each staff action on your organization is written to an audit log that your admins can read.

Infrastructure

All traffic is served over HTTPS. The connector runs on Railway, the dashboard on Vercel, and data and authentication on Supabase. Sign-in codes are emailed through Resend, and payments are handled by Stripe; MCP Tutor never sees your card number.

Reporting a problem

If you believe you have found a security issue, email support@mcptutor.com with the details. We will acknowledge it and keep you informed while we fix it.

Security — MCP Tutor