Skip to main content

​
CRM MCP Server

The TagadaPay CRM MCP Server (@tagadapay/crm-mcp) is a Model Context Protocol server that exposes a read-only view of your TagadaPay account to any MCP-compatible AI agent — Claude, Cursor, Windsurf, ChatGPT desktop, or your own. It is the same diagnostic surface that powers the in-app Auto-fix agent (the chat widget on app.tagadapay.com) and the Crisp support agent. Now you can plug it into your own AI tooling.
Read-only by design. This MCP can answer “how is my account configured?” but it cannot change anything. Mutations always go through the in-app approval flow so a merchant explicitly clicks “Apply” before any change lands. See How writes are gated below.

​
How it works

1

Get an API key

From the dashboard: Settings → API Keys → Create. Scope it to read.
2

Add the MCP server to your AI tool

One JSON snippet in .cursor/mcp.json (or your tool’s equivalent). See Quick Start.
3

Ask the agent

“Why are my payments declining?” → the agent inspects processors, webhook deliveries, domains, and answers with real data from your account.

​
What can agents read?

Thirteen tools at v1, all read-only, all backed by the Node SDK: See Tools Reference for full schemas.

​
What agents can’t do via this MCP

  • Mutate anything. No “enable processor”, no “add webhook”, no “promote funnel”. This MCP has zero write tools.
  • Cross-account access. The API key is scoped to one account. Any tool call is implicitly filtered by accountId.
  • Read PII or card data. Customer emails, card PANs, and payment instrument tokens are out of scope. Only configuration data is exposed.

​
How writes are gated

If you want an AI agent to actually fix a misconfiguration (e.g., re-enable a disabled processor), use the in-app Auto-fix agent instead. That agent runs server-side, proposes a plan, and the merchant approves it inline in the dashboard widget — every change is HMAC-signed, plan-hash-locked, and audited. The CRM MCP is the diagnostic layer. It tells you what is wrong; the in-app agent gives you the one-click fix. This separation is intentional: a misbehaving external AI agent (prompt-injected, jail-broken, whatever) can at worst make a redundant read_* call. It can never push a button on your account.

​
Pairs well with

​
Next steps

Quick Start

Get an API key and connect Cursor or Claude in 2 minutes

Tools Reference

Every tool with input/output schemas

Examples

Sample prompts for diagnosing common issues

Studio MCP

Edit your Studio templates live with an AI agent