The idea
A coding agent can work with the files and instructions you give it. But a task often depends on information outside the repository: an issue description, a support ticket, or internal documentation. Copying that information into every chat is slow and can go stale.
Model Context Protocol (MCP) is an open standard for connecting an AI application to external data sources and functions. An MCP client, such as a compatible coding agent or IDE feature, connects to an MCP server. The server describes what it can provide. MCP is a connection layer. It does not replace your task instructions, project files, or review process. GitHub describes MCP as a way to connect Copilot with external data sources and tools, and says it complements prompts, files, and instructions. Source
Plain English: MCP is a standard-shaped bridge. It can let an AI agent reach a particular outside system, but you still decide which bridge to connect and what access crosses it.
The MCP specification defines three kinds of server capability, called primitives. A prompt is a user-controlled, reusable instruction. A resource is application-controlled context, such as information a client can provide. A tool is a model-controlled function that can retrieve information or take an action. These are separate concepts. A tool is not merely extra text for the model; it is a callable function exposed by the server. Source
Think of MCP as an adapter, not as intelligence. The adapter can make an issue tracker available. It does not determine whether the agent interpreted an issue correctly or proposed a safe code change.
A small example
Example (hypothetical): Your team keeps bug reports in an issue tracker. You connect a reviewed MCP server that exposes a read-only issue lookup tool.
You ask: “Read issue DEV-42. Identify likely affected files in this repository. Propose a change, but do not edit files.”
The prompt states the goal and the boundary: propose, do not edit. The MCP tool supplies the issue data. Your repository and project instructions still provide the code context and working rules. The agent then returns a proposal that you review like any other suggested change.
This separation is useful. When something is wrong, you can ask a clearer question: was the instruction vague, was the external data incomplete, did the tool return the wrong result, or was the proposal poor?
MCP support is not identical everywhere. What a server offers and what a specific client surface supports can differ. For example, GitHub documents different MCP support across Copilot surfaces. Check the client documentation instead of assuming every prompt, resource, or tool will be available. Source
Try it
Set aside five minutes for a small, low-risk inspection.
- Pick one external system you already use, such as an issue tracker or documentation source.
- Ask whether the agent truly needs external access. If the needed information is already in the repository, no MCP connection may be needed.
- If you have an available MCP server, list the tools it exposes. Prefer one read-only lookup tool for a first test.
- Check the account, credentials, and permission scope. Grant access only to the resources required for that one task.
- Test connectivity, then ask for a summary of one known item. Compare the response with the source before using it in a coding task.
GitHub recommends reviewing third-party servers, enabling only needed tools, scoping credentials, and validating configuration. Source
Watch out
An MCP server is software and a trust boundary. Do not treat a server as safe merely because it has a familiar name or connects successfully. Review what it exposes, who operates it, which account it uses, and whether it can read, write, or trigger actions.
Use the smallest useful permission scope. Read-only access is often a safer starting point than write access. GitHub also warns that its Copilot cloud agent can use configured MCP tools autonomously, without approval before every call. That behavior is specific to that product surface, but it illustrates why tool scope matters. Source
Finally, access does not equal correctness. MCP can supply current external information or callable functions. It cannot guarantee that an agent selects the right tool, understands its output, or makes the right engineering decision. Review results and planned changes before accepting them.