AI Agents

MCP connectors explained: what to check before you connect

GitHub just made MCP connectors one-click in Copilot. Here's what an MCP connector is, what it can reach after you sign in, and what to check first.

S

SelfAgentic Team

7 min read
Blue and grey network cables plugged into some of the numbered ports on a patch panel.

On October 28, GitHub started rolling out one-click connectors for Copilot: you sign in once to a service like Slack, Figma, Notion or Sentry, and Copilot can use it from the app, the CLI, VS Code and your phone. Each of those connectors is an MCP server, built on an open standard that any AI app can adopt.

MCP connectors are a standard plug that lets an AI agent read from and act inside another app. Plugging in is now the easy part. You still have to decide which agent gets to use each connection, for what, and with whose approval, because the protocol leaves those choices to whoever builds the app you're using.

What is an MCP server, in plain English?

MCP stands for Model Context Protocol. The official MCP specification describes it as an open protocol for connecting AI applications to outside data and tools. Without a shared standard, every AI app needs custom wiring for every service. With one, a service can publish a single MCP server, and any app that speaks MCP can use it.

The spec has three roles. The host is the AI app you work in, such as a coding assistant or an agent platform. The client is the connector inside that host. The server sits on the service's side and can offer tools, which the spec defines as "functions for the AI model to execute." Sending a Slack message, creating a Notion page and searching an error log are all tools.

When a product says "connector," it usually means a ready-made MCP server plus the sign-in flow, so there's nothing for you to install. GitHub launched with 26 supported connectors, including Airtable, Canva, Linear, Notion and Slack.

The protocol itself was revised this summer. Its 2026-07-28 release made the core stateless and brought sign-in closer to OAuth 2.0 and OpenID Connect practice. OAuth is the standard behind most "Sign in with..." buttons, so connecting an MCP server now feels much like connecting any other app to your Google or Slack account.

Connecting a tool is not the same as authorising an agent

GitHub's connector documentation says Copilot "can use tools from a connected service when they are relevant to your request, even if you do not mention the service by name." That's convenient. It also means that once a service is connected, the agent decides when to reach into it.

One-click setup makes it easy to blur two questions. Connecting answers "can this app reach our Slack?" Authorising answers "which agent may post in which channel, and does a person check it first?"

The MCP spec puts the second question on the app. It says hosts "must obtain explicit user consent before invoking any tool," and also that "MCP itself cannot enforce these security principles at the protocol level." So whether you see a prompt for every action, once per session or never depends on the product.

GitHub's docs, for example, don't say what permissions each connector receives or whether individual tool calls need approval, so you have to check each service yourself, on the consent screen and in that service's admin settings.

We kept these as separate steps in SelfAgentic for this reason. Connecting a tool doesn't let every agent use it: each agent needs its own grant for each integration, and publishing actions wait for approval by default. The details are on our security page.

Are MCP connectors safe? Five checks before you connect

The protocol's own MCP security best practices list "using wildcard or omnibus scopes" as a common mistake. They recommend starting with a small set of low-risk read scopes and asking for more only when a privileged action is first attempted. Not every server works that way, so run through these before you click Connect:

  1. Read the scopes on the consent screen. A scope is one thing the login is allowed to do, such as "read channels" or "write messages." If a connector for a summary job asks for write access to everything, decline or look for a narrower option.
  2. Decide which agent gets the connection. When every agent shares one login, a mistake by any of them reaches everything that login can touch. Give each connection only to the agent whose job needs it.
  3. Treat reading and writing differently. Reading a Slack thread and posting to a customer channel carry very different risks, so put a human approval in front of anything that leaves your company.
  4. Don't rely on a tool's description of itself. The spec says tool descriptions "should be considered untrusted, unless obtained from a trusted server." A tool labelled read-only is making a claim, and the scopes are what limit it.
  5. Know where the off switch is. Find out how to disconnect the service and whether that removes every agent's access at once.

A worked example: a weekly feedback digest

Say a four-person startup wants a Monday digest of customer feedback. Comments arrive in two Slack channels and a shared Google Sheet, and the founder wants a one-page summary posted to the team channel.

A careful setup looks like this. A research agent gets read access to those two channels and the sheet, and nothing else. A writing agent can create documents in one "Digests" folder in Google Drive. Neither can post in Slack by itself. The team-channel post is drafted overnight and held until the founder approves it on Monday morning.

This is manager and specialist delegation: a lead agent splits the job, each specialist gets only the access its part needs, and the step that reaches other people waits for a person. SelfAgentic structures delegation this way, and you can copy the pattern on any platform that lets you set access per agent.

The cost is setup time, since you'll go through scopes and grants for each agent instead of clicking one Connect button. The payoff shows up on a bad day. If the research agent reads a message with hidden instructions in it (an attack called prompt injection), the worst it can do is write a misleading summary, and you see that summary before anyone else does.

FAQ

Is MCP the same as an API?

No. An API is how one service lets software talk to it. MCP is a shared layer on top, so an AI app can use many services through one standard format instead of custom code for each.

Do I need to be a developer to use MCP connectors?

Not for the ready-made kind. Products like GitHub Copilot now handle the server setup and sign-in, so connecting mostly means picking a service and approving access. Building or hosting your own MCP server is still developer work.

Are MCP connectors secure?

The protocol sets rules for sign-in and consent but can't enforce them, so the app you use decides how prompts and permissions work. Check the scopes, limit which agents can use each connection, and require approval for actions that leave your company.

What's the difference between connecting a tool and giving an agent permission?

Connecting gives the app a working login to the service. Permission decides which agent may use that login and for what. Platforms that keep the two apart let you connect once and still give each agent only what its job needs.

What MCP connectors don't solve yet

  • GitHub's connectors are in public preview and "subject to change," and they don't work in GitHub Actions sessions, such as when you assign an issue to Copilot.
  • Some services need an admin to switch MCP access on first. GitHub's docs use Slack as the example, and Copilot Business or Enterprise users need an admin to enable preview features.
  • When a server doesn't say which scope an action needs, the spec tells clients to fall back to requesting every scope the server lists. The consent screen is where you can push back.
  • A connector tells you what an agent could do. To see what it did, you still need a run log of which tools each agent called.

A practical first step: pick one connection your team already has, in any AI tool, and write down which agents or assistants can use it and what its scopes allow. If the answers are "all of them" and "everything," fix that one before you add a new connector.

If you'd like a team of agents where connecting a tool, granting it to an agent and approving what goes out are three separate steps, you can try SelfAgentic.

Keep reading

Ready to build your AI team?

See how a team of agents handles real work, and checks with you before anything goes out.

Free to start. No credit card required.