MCP explained for ops teams: how your agents connect to company context
MCP is the open standard that lets AI agents use outside tools and data. Here is how it works, who controls access, and where Nysa fits, in plain language.
MCP, the Model Context Protocol, is an open standard that lets an AI agent connect to outside tools and data. A server offers a set of tools, such as "search our meetings", and a client, such as Claude or ChatGPT, calls them. You sign in once and the agent can use those tools, with whatever access your account has.
If you run operations, you do not need to build anything to benefit from it. You need to understand what you are switching on, who can see what, and what to check.
The plain-language picture
Think of an agent as a smart new colleague with no access to anything. MCP is the standard way to hand that colleague a keycard to a specific room.
Three parts matter:
- Server. The service that holds your data or does the work. It lists the tools it offers, like "find a person" or "search company knowledge".
- Client. The app where you talk to an agent, such as Claude, ChatGPT, Cursor or Codex. It connects to servers and shows the agent which tools exist.
- Tools. The individual actions a server exposes. Each has a name, a description and inputs. The agent decides when a tool would help and calls it.
Because the standard is open, a server built once works with any client that supports MCP. Your team does not need to pick one agent for everyone.
What happens when you ask a question
Say you ask your agent: "What did the customer say about the renewal last week?"
- The client already knows the server's tools, because you connected it earlier.
- The agent sees a search tool fits the question and calls it with your query.
- The server checks who you are, looks up only what you are allowed to see, and returns results.
- The agent reads the results and writes you an answer.
The agent never has your whole database. It gets the slice the tool returns for you.
Access: how permissions work
For ops teams, this is the part that matters. Most remote MCP servers use OAuth, the same "sign in with Google" flow you already know. When you connect, you sign in to the service and approve the connection. The agent then acts as you.
That has two consequences:
- Access follows the person. A well-built server returns only what the signed-in person can see, not everything in the system.
- Access can be revoked. Remove the person or the connection and the agent loses the keycard.
Questions to ask about any MCP server before you connect it:
- Does the agent act as me, or as a shared account?
- Which tools can change data, and which only read?
- Is every call logged?
- Can an admin switch it off?
Be careful with tools that can write or act. Reading company knowledge is low risk. Sending messages or editing records deserves a closer look.
Where Nysa fits
Nysa is an MCP server for your team's shared memory. Meetings, Gmail and Google Calendar become facts with sources, and Nysa serves them to your agents at https://mcp.usenysa.com/mcp, using OAuth sign-in.
What that gives you in practice:
- One brain, many agents. Claude, ChatGPT, Cursor, Codex, OpenClaw, Hermes or any agent that connects over MCP can use it.
- Per-person scope. Each agent is signed in as its teammate and sees only what that teammate can see. Raw email stays private to its owner.
- Read and write. Agents can look things up and also add or correct facts. On the Claude page, updates happen only on request.
- Audit trail. Calls are logged, and removing someone's access stops further requests.
- Limited actions. Claude cannot send emails or edit calendars through Nysa.
To see the tools in detail, read Nysa MCP tools explained.
How to connect
The steps take a few minutes.
- Open your agent's connector settings. In Claude, that is Customize, then Connectors.
- Search for Nysa, or add a custom connector with the server URL above.
- Choose Connect and sign in with the Google account you use for Nysa.
- Start a new conversation. The tools load automatically.
In a Claude workspace managed by an owner, the owner may need to enable custom connectors first. Step-by-step guides: Connect Claude and Connect ChatGPT. For the full workflow, see connect ChatGPT to company meetings, email and calendar.
A short checklist for ops
- List which agents your team uses and which MCP servers each one has connected.
- Prefer servers where the agent acts as the signed-in person.
- Separate read-only tools from tools that change things.
- Confirm there is an audit log and a way to revoke access.
- Review connections when someone leaves.
MCP is simple at its core. The work for ops is the same as with any integration: know what is connected, who it acts as, and how to turn it off.
FAQ
Is MCP only for developers?
No. Developers build MCP servers, but connecting one is usually a settings screen and a sign-in. Ops and team leads can set it up without writing code.
Is MCP secure?
The standard itself is a way to connect; security depends on the server. Look for OAuth sign-in, access that follows the signed-in person, audit logs and a way to revoke. Review any server you connect, especially ones that can write or take actions.
Do I need one MCP server per agent?
No. That is the point of an open standard. One server, like Nysa's, works with any MCP client, so Claude, ChatGPT and others can share the same company context.