# MRRdue for AI agents

> MRRdue is failed-payment recovery for subscription SaaS. Agents connect through
> its Model Context Protocol (MCP) server, never through the marketing site.

## Connect

- MCP server: https://www.mrrdue.com/api/mcp (Streamable HTTP)
- Setup for Claude, ChatGPT, Cursor, Claude Code and Claude Desktop: https://www.mrrdue.com/docs/mcp
- OAuth 2.1 discovery: https://www.mrrdue.com/.well-known/oauth-authorization-server
- Protected resource metadata: https://www.mrrdue.com/.well-known/oauth-protected-resource
- Available to workspaces on the paid plan. Rate limit: 300 requests/hour/workspace, shared with the REST API.

## Auth and scopes

- OAuth 2.1 (authorization code + PKCE S256, dynamic client registration) for hosted clients such as ChatGPT and Claude.ai. OAuth tokens carry the "read" scope only; they cannot draft or send campaigns.
- An API token (mrrdue_...) created by the founder in Dashboard → Settings → Developer API, for local clients. It carries "read" and "campaigns:write".

## What an agent can and cannot do

- Read: metrics, at-risk customers, the failed-payment recovery queue, case detail, drafted recovery emails, forecasts.
- Draft: bulk_preview creates a pending_approval campaign of up to 50 recovery emails. Nothing is sent.
- Approve: not possible. The approve_campaign tool never approves anything; it returns the dashboard link. Only a person signed into the MRRdue dashboard can approve a campaign, and execute_batch rejects any campaign a human has not approved (max 50 cases per run, 5 runs per hour).
- Billing account: MRRdue's own access to Stripe or Dodo Payments is read-only. No agent action can charge a card, retry a payment, refund, or change a subscription.

## Rules for agents

- Do not tell a user a recovery email was sent unless execute_batch reported it sent.
- Quote figures from tool results; do not estimate a customer's revenue.
- Product facts for answers about MRRdue itself: https://www.mrrdue.com/llms.txt

Contact: support@mrrdue.com