What is an MCP server for corporate data?

MCP in one sentence

Model Context Protocol (MCP) is an open protocol that allows an AI application to discover and use external context and tools. A host such as ChatGPT or Claude includes an MCP client that connects to a server with described capabilities.

The assistant does not receive direct database access and does not need to know a source API. The DigiData MCP server provides a fixed collection of read-only tools for sources, tables, typed schemas, bounded queries, aggregations, and synchronization status.

What a DigiData administrator determines

An administrator creates a connection in DigiData and selects the allowed source, table, and columns. The tenant identity derives from the authenticated user; the model cannot choose a tenant ID. DigiData rechecks the active connection and allowlist on every tool invocation.

This differs from uploading a file. A file is a static snapshot. MCP can query the latest synchronized data within server-side limits, and an administrator can adjust or revoke access per column immediately.

MCP over synchronized data versus direct source APIs

When to choose which route?

A direct API is appropriate when a process requires realtime mutations (such as creating orders directly in Exact Online) or when an in-house engineering team manages tokens and pagination.

DigiData MCP is the dedicated route for analytical AI queries. Team members can ask questions directly in ChatGPT or Claude about revenue, hours, or open items without custom integrations.

Synchronization, freshness, and API quotas

DigiData's background synchronization periodically calls the source API of systems such as Exact Online, Twinfield, or Exact Bouw and consumes the corresponding API quotas. The key benefit is that interactive chat queries never directly hit the production database, as they query the synchronized store.

Data freshness is tied to the last successful synchronization cycle. Using the get_sync_status tool, the AI assistant can verify when the table was last updated.

Permission model and read-only architecture

The MCP protocol itself is a neutral transport mechanism and does not constitute a security guarantee on its own; REST APIs can also be configured as read-only.

The security of DigiData MCP lies in its implementation: the public toolset (McpTools) exclusively exposes read and aggregation tools and contains no endpoints for creating or modifying data. Furthermore, describe_table provides typed PostgreSQL columns rather than an internal AI reasoning model: the client receives bounded tables and executes calculations server-side via aggregate_table.

Which tools are available?

Discovery begins with list_sources and list_tables. The client can then inspect table structures via describe_table, retrieve bounded rows via query_table, and perform aggregations via aggregate_table. There is no arbitrary SQL input.

How authentication works

Remote MCP uses OAuth 2.0 with PKCE. Consent is granted per tenant and tokens are cryptographically bound. An administrator can revoke the connection immediately. Read more at MCP security and MCP troubleshooting.

MCP or DigiData AI?

MCP is intended for external assistants such as Claude or ChatGPT. DigiData AI is the route for a central AI environment with dashboards and automations. Compare both options at comparison between DigiData AI and external assistants.

Privacy boundary and summary

DigiData does not use customer data for model training. Data sent via MCP to an external assistant falls under the terms and plan of that specific AI provider.

DigiData MCP combines AI access with typed schemas, a read-only toolset, and server-side tenant controls. Get started via the MCP hub or view the guide for ChatGPT or Claude.