← All posts
AI   Nov 19, 2026 · 6 min read · by , Co-Founder

What an MCP Server Gives an AI Agent, and What It Leaves Out

In short

An MCP server gives an AI agent a standard way to call a tool: to search Jira or post in Slack. It supplies access. It tells the agent nothing about which goal a ticket serves or whether a project is behind, unless the server sits on a system that already holds that context.

Anthropic published the Model Context Protocol in November 2024, and within a year OpenAI and Google had adopted it. Few standards spread that fast. The reason is plain enough: before MCP, connecting an agent to a company tool meant custom code for every pairing of agent and tool, and after it, a tool vendor writes one server and every agent can use it. If your company runs agents today, they almost certainly reach Jira or Slack this way.

It is tempting to conclude that the context problem is thereby solved. The agent can reach everything, so it knows everything. That conclusion mistakes a connection for knowledge.

What does an MCP server provide?

An MCP server describes a set of tools to the agent, each with a description and its inputs. The agent decides which tool to call and the server carries out the call, with the user's sign-in and permissions applied. A Jira server offers a ticket search. A Slack server offers message history. The protocol standardizes how the agent asks, and that is a large practical gain, because your access rules now live in one place per tool.

However, each server returns only what its own tool holds. The Jira server returns tickets. It has no idea that ticket PLAT-212 belongs to the project your board is most worried about, because Jira holds no record of your board's worries. The Slack server returns messages and cannot say that one of them announced a blocker on that same project. Each server knows its own tool and nothing of the plan that runs across them.

What does the agent still have to work out?

So the agent does the joining itself. Asked whether the onboarding initiative is on track, it searches Jira and then Slack, pulls back several thousand words of raw material, and tries to reconstruct which pieces belong to the initiative and what they add up to.

Sometimes it does this well. It also does it differently each time, since the search terms it picks and the results it gets vary from run to run, a problem we describe in Why AI Agents Get Your Company Context Wrong.

There is a cost in volume as well. Every connected server adds its tool descriptions to the agent's context window, and every call adds its results. Connect ten servers and a good share of the window is spent before the agent has thought about your question. More connections, past a point, make the agent slower and less accurate.

What should stand behind an MCP server?

The useful question about any MCP server is therefore what stands behind it. A server in front of a ticket system gives the agent tickets. A server in front of an AI context layer gives the agent the joined picture, with the work of assembling it already done and done the same way for everyone.

The Vindaris MCP server is of the second kind. Behind it sits the Work Graph, which maps the strategy to the goals, KPIs, projects and tasks under it and analyzes work activity across the tools your teams use, such as Jira, Slack, Microsoft 365 and meeting transcripts.

One call to get_work_graph_alerts returns the items that need attention, each with its reasons, weeks before a review would raise them. One call to get_goal returns a goal with its links. The agent spends its context window on your question.

Access is bounded in the usual MCP manner and a little beyond it. An admin sets each resource (strategy, goal, KPI, project, task) to off, read, or read and write, with off as the default, and changes an agent proposes to the plan can wait for a person's approval. The setup for each client is on the AI agents page, and a step-by-step version is in How to Give an AI Agent Context on Company Goals.

Do more MCP servers make an agent better?

None of this argues against MCP, of course. The protocol did its job, which was to make access cheap. Nonetheless, cheap access has moved the scarce thing elsewhere. Two years ago the hard part was getting an agent into your tools at all. Today any agent can get in, and what distinguishes one setup from another is whether something on the far side of the connection already knows how your tickets relate to your goals.

When you next review the list of servers your agents are connected to, which of them could tell an agent what your company is trying to achieve this year?

FAQ

What is an MCP server? An MCP server is a program that offers tools to an AI agent over the Model Context Protocol, a standard Anthropic published in November 2024. The agent calls a tool and the server carries out the request with the user's permissions.

Does connecting Jira and Slack over MCP give an agent company context? It gives the agent access to tickets and messages. The agent still has to work out on every run which of them belong to which goal, and it will not always reach the same result.

How many MCP servers should an agent have? As few as answer the questions you ask. Every server adds tool descriptions to the context window, so a long list leaves less room for the task.

What does the Vindaris MCP server return? A joined view of the plan: goals with their linked KPIs, projects and tasks, and Work Graph alerts with the reasons behind them. It runs at app.vindaris.com/api/mcp with OAuth sign-in.