The Model Context Protocol (MCP) is the open standard that connects AI assistants to your applications. Instead of building a separate integration for every assistant, you expose your domain logic once as an MCP server – and your users operate your application from the AI client they already work in: Claude, ChatGPT, or tomorrow's internal tool.
We build MCP servers as thin adapters over a shared tool layer: the same domain functions serve the in-app chat assistant, external AI clients via MCP, and any future agent. The logic lives in one place – the channels are interchangeable.
Security is not an afterthought but the architecture: full OAuth2 with dynamic client registration and a consent screen, tool categories as scopes, role-based guards on every call, and audit logging for every tool execution. What an assistant isn't allowed to do, it never even gets to see.
And because we believe in our own tools: this website itself runs as an MCP server. Pages, articles, and media of this site can be managed straight from an AI assistant – through the same architecture we build for our clients.
An assistant can never do more than the logged-in user
Every tool call passes two checks: the OAuth scope caps what the token may do – the user's role caps what the person may do. Only when both agree does the call run.
Tool categories are the OAuth scopes. Ungranted tools are hidden entirely rather than refused – the assistant only sees what it is allowed to use.
Write actions are never executed directly but staged as draft proposals a human reviews and approves in the UI – and every call lands in the audit log.
Three systems we opened up as MCP servers – from a member portal to this very website.
For a German film-industry association we expose the member portal's domain services once as a shared tool layer – served through two channels: a chat assistant inside the member area, and an MCP server through which members operate the portal from their own AI assistant. The MCP channel speaks full OAuth2 with dynamic client registration and a consent screen; every call passes the double guard of scope and role.
A licensing and royalty platform for a global media group must resolve every incoming title to the house product identifier before revenue can be reported. The platform's deterministic mapping engine is exposed as an OAuth2-protected MCP server: a business user pastes a distribution partner's title list into their AI client and gets back a structured verdict per row – the identifier with explicit certainty, or an ambiguous/no-match status. The model never decides a mapping: the agent asks, the engine decides.
The site you are reading right now runs as an MCP server itself – secured by an OAuth2 firewall. Creating pages, maintaining content, browsing media: our editorial team works with the CMS straight from an AI assistant. The same architecture we build for client projects carries our own web presence.
Why MCP is more than an interface
The real win of MCP is architectural: the chat assistant, external AI clients, and future agents become thin adapters over a single, role-guarded tool layer. New domain functions are automatically available in every channel, permissions are enforced in one place, and answers cannot drift between channels – because the same engine answers everywhere.
