MCP and skills at enterprise scale

Why this conversation

Rolling MCP servers out to a team of developers is one problem. Rolling them out to a company of a hundred thousand people, almost none of whom know what MCP is, is a different one. Someone has to decide who gets which tools, keep them secure and audited, and keep them running. At the close of AGNTCon + MCPCon Europe in Amsterdam, on 18 September, the host spoke with Patrick Eden, CEO of Palma AI, which provides governance for MCP and skills at scale. The rest of the week is in AGNTCon + MCPCon Europe 2026: who we met.

Agents get simpler

Asked why he came, Patrick calls the conference one of the places to see what the community is doing at the frontier of MCP agents. What he has enjoyed watching over the last year and a half is a great simplification. Agents used to be complicated to build, with custom frameworks for very vertical use cases. Now they are collapsing into something much simpler: “Harness plus LLM is almost enough, plus MCP for the connection to a system”. The frontier vendors, he says, are all going through that same evolution.

One connector for everyone

In a big enterprise of, say, 100,000 people, he says, nearly all of them are not technical. They do not know what MCP or a skill is, and probably will not care. They want it to work. MCP and skills are superpowers for agents, and the goal is to enable them for everyone in the company. Palma does that with a single connector: everyone gets one MCP, and central IT manages everything behind it, so the CEO, someone in accounting and someone in sales each get the right tools for the job. Behind that connector sit the policies, audits, security scanning and runtime skills.

That is how Palma sells to companies with as many as 600,000 employees. Scale, he says, changes everything. Beyond reliability, such companies need approvals and distribution, and the product has to work with many distributed teams: compliance, governance and security each want their own dashboard and functionality.

Critical middleware, on-prem

The host suggested a service like this becomes mission critical, where an hour of downtime has a big impact. Patrick agreed, adding that where there is value there is criticality. Palma sees its role as critical middleware and is careful about it. It also ships the product on-prem, and the largest customers tend to run it that way. In his experience, downtime is rarely caused by Palma itself, and more often by something like a failed Kubernetes deployment. It can be deployed wherever a customer wants, including in high availability setups. He acknowledges that enterprises are not naturally eager to trust startups with highly critical software, and says Palma works hard to earn that confidence.

Collaborating on skills

What excites him most is a skill collaboration system Palma is working on. MCP access at scale is complicated enough, but skills have become a way to distribute very specific knowledge: playbooks and knowledge an agent loads dynamically based on what it is doing. The questions at scale are who owns a skill, how 10,000 people work with one and suggest changes, and how the agent consolidates those changes. There is also a craft to writing a skill, with the right amount of context, and the need to evaluate it: who uses it, how it is distributed, whether it is scoped to the right MCP servers, and whether everyone who gets the skill has the MCP access it needs.

Skills improve agents even when the underlying assistant is not the strongest. He mentions feedback from executives using Palma with Microsoft Copilot on their iPads, who found it could suddenly run processes and pull context from their systems.

How to try it

Palma does not offer a freemium model. People get in touch through palma.ai, and the team usually runs proofs of concept.