Convertr MCP: Data & Security Overview Follow

Purpose

This document explains what data Convertr MCP can access, how authentication and authorization work, and what security controls are in place. It's intended to answer the questions customers ask before connecting an AI tool to Convertr.

1. What Data Is Accessible

Convertr MCP v1 exposes the same read-only reporting data as the in-platform Reporting Assistant, through the existing Convertr Reporting Agent — MCP is a new access route to that same agent, not a new or expanded data surface. Supported query categories:

Campaign & delivery data

Lead data fields

Example questions supported (same phrasing patterns as the Reporting Assistant): "How many leads were delivered yesterday?", "Show lead delivery for the last 7 days", "Show valid leads for Campaign X last week", "Show billable lead counts by company size, job role and industry", and natural follow-ups like "Can you break that down by Publisher?"

Data freshness: query results reflect data refreshed on the same ~20-minute cycle as the Reporting Assistant — this is a refresh window, not a live/real-time feed.

2. What Is NOT Accessible / Out of Scope

Carried over directly from the Reporting Assistant's own limitations, since MCP calls the same underlying agent:

• No write access of any kind. The agent cannot create, edit, pause, or delete campaigns, publishers, or any other Convertr object — under any phrasing of the request.
• No cross-tenant or cross-account data. A user cannot access another tenant's or another user's data through the connector.
• No configuration or setup data — value transforms, agencies, forms, batch job schedules, campaign metadata, users, and audit logs are all out of scope.
• No explanatory or troubleshooting answers — the assistant reports what happened, it doesn't explain why a configuration behaves a certain way or diagnose issues.
• No system internals exposed — no SQL, underlying data sources, or processing logic is ever surfaced in a response.
• No real-time streaming — see the ~20-minute refresh cycle noted in Section 1.
• MCP responses are text-only. The Convertr MCP server returns natural-language/text answers to queries — it does not itself generate or serve files (CSV, charts, decks, etc.). Any file or visual output (a chart, spreadsheet, or summary document) is produced downstream by the connected AI agent — e.g. Claude or ChatGPT turning the returned text/data into a chart or table within that conversation — not by Convertr. There are no deep links back into the Convertr platform from the external AI tool.
• No persisted conversation history in-platform — session context lives only for the duration of the conversation.

3. Authentication & Authorization

• Convertr MCP uses OAuth 2.0 with PKCE, JWT/JWKS, and Dynamic Client Registration.
• Authentication happens entirely through a Convertr-hosted login and consent screen — the AI platform (Claude, ChatGPT, Teams & Slack) never sees or stores the user's Convertr credentials.
• The consent screen explicitly tells the user what the connection will and won't do (e.g., "will be able to use AI Reporting to query reporting data on your behalf; will not be able to edit, delete, or change any data, or access other users' accounts or data").
• No data is returned until a valid, unexpired token tied to a real Convertr account is presented — unauthenticated requests are rejected at the server.
• SSO is supported where already enabled on the customer's Convertr account.

4. Data Scoping & Multi-Tenancy

The MCP server enforces the same data permissions as the in-platform Reporting Agent — no new data is made accessible externally that wouldn't already be visible to that user inside Convertr. Tenant isolation is enforced at the point of query, not layered on separately.

5. Token Handling & Session Security

• Access tokens are scoped to the authenticated user's existing permissions.
• Tokens are intended to be short-lived and are not exposed in logs, URLs, or response payloads.
• A session begins on the first authenticated query and ends when it expires or the user disconnects.

• All traffic between the AI platform, the Convertr MCP server, and Convertr's backend is encrypted in transit.

6. Network & Infrastructure

• The Convertr MCP server is reachable at https://<yourdomain>.cvtr.io/ai/mcp.
• The server must be reachable over the public internet from each AI vendor's infrastructure — it cannot sit behind a VPN-only or IP-restricted network for hosted AI clients 

• The endpoint is not intended to be publicly usable without authentication; rate limiting and request validation are applied at the server layer.

7. Common Questions