About Us

Blinqx offers AI workflow platforms that empower service professionals to grow and succeed.

Sectors

Blinqx develops AI workflow platforms for financial and business service providers in selected sectors.

Insights

Stay up to date on what's going on at Blinqx: awards, acquisitions, knowledge, cases. You can find it here!

Search

MCP: What Platform Builders Need to Decide Now

Modified on

More and more customers are letting agents do the work in our software, and those agents communicate with your platform via an MCP server. This makes MCP the new front door to your product. Whoever consciously decides now which features to make available, how to handle billing, and how to maintain control will determine whether agents will eventually work through your platform or around it.

What is MCP, and why does it affect platform builders?

MCP, the Model Context Protocol, is an open standard that allows an AI model to retrieve structured data and perform actions in other software. An MCP server defines which tools an agent is allowed to use and what context it receives. For us as a provider, this means: a second user alongside humans, with its own requirements for your product.

Anthropic introduced MCP in late 2024, and the standard was quickly adopted by model providers and development tools. This is relevant for platform builders in the business and financial services sectors, because our clients work with sensitive files, deadlines, and accountability requirements. An agent who reschedules a deadline in a practice management system or posts a journal entry in an accounting package is performing work with legal and financial implications.

A classic REST API was designed for developers building integrations. An MCP server is designed for a model that determines which step to take on its own. That difference changes the questions you need to answer as a CTO or CPO.

Which functions do you make available to agents?

Start with tasks that have a clear outcome and a verifiable step, such as searching, summarizing, and preparing drafts. Only make actions that finalize something available later, and always with confirmation from the professional. This way, autonomy grows alongside trust.

We use three layers when designing an MCP server:

  1. Reading. An agent may access files, client records, time logs, or invoices within the scope of the user’s permissions. This is where the greatest value lies, and the risk is limited.
  2. Preparation. The agent creates a draft: a letter, a booking proposal, or a timesheet. The professional approves or edits it.
  3. Execution. The agent takes a final action themselves, such as sending an invoice or changing a due date. You should only grant access to this level for well-defined tasks, with logging and a clear owner.

The pitfall is publishing your entire API one-to-one as MCP tools. A model would then be presented with hundreds of options without any context about when to use which tool. A better approach is to use a smaller set of tools that aligns with how a lawyer, accountant, or consultant thinks. A tool like “find all outstanding deadlines in this case” works better than three separate endpoints.

How do you maintain control over what an external agent does?

Control starts with identity and permissions: an agent always acts on behalf of an identifiable user and never with more permissions than that user. In addition, you log every request so that you can reconstruct afterward who did what. In regulated sectors, that audit trail is the foundation of trust.

This aligns with what we call “compliance-by-design ”: you build the controls into the product from the very beginning. Specifically, for an MCP server, this means:

  • Authorization on a per-user basis. The agent inherits the permissions of the logged-in professional. No generic service keys with access to all customer data.
  • Scopes per tool. A client can specify, on an organization-by-organization basis, which tools an agent is allowed to use. An office that only grants agents read-only access must be able to configure this setting.
  • Human-readable logging. Record which tool was called, with which parameters, and what the result was. A compliance officer must be able to track this without being a developer.
  • Data location. Know which models access your data through the client’s agent and where they are running. Under the GDPR and the AI Act, the client remains responsible, and we, as the provider, must enable that responsibility.

Third-party agents expand the playing field. A customer can connect their own agent or an agent from another vendor. Your MCP server must therefore also be able to handle agents that you did not build yourself. Treat every call as input that you validate.

How do you bill when an agent uses your product?

Agents are best billed based on usage. With agents, the value shifts from the number of users to the work being done, and a per-seat model does not reflect that shift. Usage-based billing ties the price to the work that actually takes place on your platform.

In the traditional SaaS world, a firm paid per user. When an agent takes on a large portion of the routine work, the number of actions per employee in your interface decreases, while the value of your platform increases. With a seat-based model, your revenue would actually decrease. That’s why at Blinqx, we opt for usage-based billing for agents, with clear cost items for usage, implementation, training, and governance.

Three decisions you need to make upfront:

  • What is the unit of measurement? A call, a completed task, or a processed case. Choose a unit that your client understands and can relate to their own work.
  • How do you make it predictable? Clients want a budget. Insight into usage per team and per period helps. In eBlinqx WorQX, we therefore make the use of AI credits visible to users and managers.
  • What’s included in your product? Determine which MCP tools are part of the existing package and which tools you’ll offer as a separate usage component.

Where do you start if you don’t have an MCP server yet?

Start with a small, easy-to-understand MCP server designed for a single specific task in a single sector, and test it with a few customers. You’ll quickly learn which tools a model actually uses and where it runs into problems. Don’t expand until logging, permissions, and billing are working properly.

A practical sequence:

  1. Choose one task that involves a lot of repetition and has a clear outcome, such as retrieving file information or preparing a draft invoice.
  2. Design two to five tools in the professional’s own language, with clear descriptions. The model reads these descriptions, so write them as if for a smart colleague with no prior knowledge.
  3. Build in authorization and logging from day one.
  4. Track usage for each tool. Remove any tools that no one uses.
  5. Discuss the revenue model with a few customers before rolling it out on a large scale.

On Tech TalQX, the podcast about AI in regulated industries, we regularly delve deeper into these kinds of design choices. We previously wrote about agent orchestration in regulated industries, focusing on how to get multiple agents to work together.

Conclusion

An MCP server is to agents what an interface is to people. Platform builders who currently determine which features to make available, how to configure permissions and logging, and how to handle billing retain control over their product. Start small, build in the controls, and let usage dictate how quickly you scale.

What is an MCP server?

An MCP server is an interface based on the Model Context Protocol that allows an AI model to retrieve data and perform actions within a software platform. The server describes which tools are available and what context the model receives. This enables an agent to perform tasks independently within your product.

What is the difference between an MCP server and an API?

An API is designed for developers who are building a fixed integration. An MCP server is designed for an AI model that decides for itself which step to take. That is why an MCP server contains fewer, better-documented tools that align with the user’s language.

Which features should you make available first via MCP?

Start with read-only functions, such as retrieving files or invoices, and with functions that prepare a draft. Actions that finalize something should only be enabled later, after confirmation by a professional. This way, you build trust before granting the agent more autonomy.

How do you stay compliant when third-party agents use your platform?

Always have an agent act on behalf of an identifiable user with that user’s permissions, and log every request in a readable audit trail. Give customers control, on an organization-by-organization basis, over which tools an agent is allowed to use. This ensures that the customer remains responsible, and you can demonstrably support that responsibility.

What revenue model works best for agents who work through MCP?

Usage-based billing is the best approach, because an agent’s value lies in the work they do. You link the price to the work performed through your platform, such as tasks or processed cases. Combine that with insights into usage, so customers can manage their budget.

Related articles