The EU AI Act and MCP: What Changes in August 2026

The compliance clock is ticking for anyone building or deploying AI agents in Europe. April 2026.

The Model Context Protocol is how AI agents connect to tools. An MCP server gives an agent access to your filesystem, your database, your browser, your API keys. The EU AI Act does not mention MCP by name, but it does not need to. MCP servers are the connective tissue between AI models and real-world actions, and that puts them squarely in scope.

I have been thinking about what the August deadline means for MCP developers and for companies that deploy agent-based systems in Europe. Here is what I have found.

What the AI Act requires

The relevant parts for MCP are scattered across several articles, but they boil down to four requirements:

1. Risk classification

AI systems are classified as minimal, limited, high, or unacceptable risk. Most MCP-based agent setups will fall into the "limited" or "high" category depending on what they do. An agent that reads and summarizes documents is probably limited risk. An agent that makes hiring decisions, processes financial transactions, or controls critical infrastructure is high risk.

High-risk AI systems must meet extensive requirements: risk management, data governance, technical documentation, human oversight, accuracy and robustness guarantees, and cybersecurity measures.

2. Transparency obligations

Any AI system that interacts with people must disclose that it is an AI. This applies to chatbots, content generators, and agents that send emails or make phone calls. If your MCP-powered agent contacts customers, it must identify itself as artificial.

For MCP server developers, this means thinking about whether your server enables interactions that need transparency disclosures, and building in the hooks to support them.

3. GPAI model obligations

Providers of general-purpose AI models (the foundation model companies) must maintain technical documentation, comply with EU copyright law, and publish training data summaries. If a GPAI model is classified as posing "systemic risk," additional requirements kick in: adversarial testing, incident monitoring, and cybersecurity protections.

This affects MCP indirectly. If the model behind your agent has compliance obligations, the tools you connect to it (via MCP) become part of the compliance surface.

4. Supply chain responsibility

This is the one that matters most for MCP. The AI Act creates obligations along the entire AI supply chain. If you deploy a high-risk AI system, you are responsible for ensuring that all components, including third-party MCP servers, meet the requirements. You cannot outsource compliance by using someone else's server.

What this means for MCP server developers

If you maintain an MCP server that is used in the EU, you should be thinking about three things.

Documentation. The AI Act requires technical documentation for AI systems. For MCP servers, this means clearly documenting what permissions your server requests, what data it accesses, what it does with that data, and what risks it introduces. A README is not enough. You need structured documentation that an enterprise compliance team can review.

Security baseline. High-risk AI systems must demonstrate cybersecurity resilience. For MCP servers, that means code safety (no eval, no injection vectors), supply chain security (locked dependencies, no suspicious packages), and permission minimization. Running an audit against your own server before your users' compliance teams do is a good idea.

Transparency hooks. If your server enables agent-to-human interactions, consider building in metadata that downstream systems can use for transparency disclosures.

What this means for companies deploying agents

If your company uses MCP-based agents in the EU, your compliance obligations include the MCP servers those agents connect to. Before August, you should:

• Inventory your MCP servers. Know every server in every agent's config. Document what each one does and what permissions it has.
• Classify your AI systems. Determine the risk level of each agent deployment. An internal summarization tool is different from a customer-facing decision system.
• Audit third-party servers. You are responsible for the security and compliance of servers you did not write. Run audits. Review permissions. Check supply chains.
• Establish human oversight. High-risk systems must have human oversight mechanisms. Design your agent workflows so a human can intervene, review, and override.
• Document everything. The AI Act is documentation-heavy. Keep records of your risk assessments, audits, incident reports, and oversight procedures.

The enforcement reality

The EU AI Office started operating in early 2025, and national authorities are setting up enforcement structures now. Fines are significant: up to 35 million EUR or 7% of global annual turnover for the worst violations. For prohibited AI practices, it is even higher.

That said, enforcement will likely start with the most visible cases: large companies, high-risk deployments, consumer-facing systems. A small team running an internal agent with MCP servers is unlikely to be the first target. But the obligations still apply, and enterprise customers will increasingly require compliance evidence from their vendors and partners.

What to do now

August 2 is four months away. If you have not started thinking about AI Act compliance for your MCP deployments, start now. The minimum viable steps:

1. List every MCP server in your organization's agent configs
2. Run a security audit on each one (you can use audit.pyfio.com for public repos)
3. Classify your AI systems by risk level
4. Start building the documentation the Act requires
5. Talk to your legal team about what "high-risk" means for your specific use cases

The companies that treat MCP security as a compliance checkbox will struggle. The ones that treat it as engineering practice will be fine.