How to add an MCP server to GitHub Copilot

• 01
  Confirm the policy is enabled first

  On Copilot Business or Enterprise, MCP is governed by an organisation policy that GitHub documents as disabled by default. If it is off, a perfectly correct config file will do nothing, and you will spend an afternoon debugging JSON.

• 02
  Create `.vscode/mcp.json` with the `servers` key

  This is the single most common failure. `mcpServers` is the key three other surfaces use and it is not the key here — a file using it is syntactically valid and semantically empty.

• 03
  Declare secrets as inputs, not literals

  Use an `inputs` entry with `type: "promptString"` and reference it as `${input:id}`. VS Code prompts once and stores the value outside the committed file, which is the difference between a shareable config and a leaked credential.

• 04
  Pin stdio server versions

  As everywhere: a launcher with no version resolves to whatever is published at start time, so approval and execution drift apart silently.

• Open the workspace and check that the server starts — a server declared under the wrong key produces no error, just no tools, which is why this check comes before any other debugging.
• Confirm the tools appear in Copilot Chat's tool list, and read their descriptions rather than trusting the names.
• If nothing appears and the file is correct, check the organisation policy before checking anything else. On Business and Enterprise it is the gate, and it is off by default.

• ·
  Treat `.vscode/mcp.json` as a CODEOWNERS-protected path

  It grants agent capability to every contributor who opens the repository. If your repository protects CI configuration from unreviewed change, this file belongs in the same rule.

• ·
  Read tool descriptions before approving the pull request

  The descriptions are model-facing instructions. A reviewer skimming for valid JSON will not notice a description that tells the agent to do something it was never asked to do.

• ·
  Check whether the same server is already in another tool

  Copilot's format differs, but the underlying server does not. If it is already reviewed for the terminal agent, reuse that review rather than starting a second one — and if it is not reviewed anywhere, this is the moment.

• ·
  Do not rely on the org policy as an inventory

  The policy is a permission gate. It says MCP may be used; it does not tell you which servers exist in which repositories, and it does not cover engineers outside the subscription it applies to.

• ·
  Organisation MCP policy

  GitHub documents an MCP policy for Copilot that an organisation or enterprise configures, described as disabled by default — so MCP is opt-in at the org level rather than something to switch off after the fact.

• ·
  Subscription scope of the policy

  The policy applies to users covered by a Copilot Business or Enterprise subscription from an organisation that configures it. Engineers on individual paid plans are outside its reach, which is worth knowing before you treat the switch as complete coverage.

• ·
  Content exclusions

  Separate from MCP, GitHub documents content exclusion configuration for Copilot, which constrains what the assistant may read. It is adjacent governance rather than an MCP control, and the two are often confused in policy documents.

The policy switch and the config file sit at opposite ends of the problem and neither closes it. One says whether MCP is permitted across an organisation; the other declares servers per repository and per machine. Nothing in between reports which servers are actually configured across your repositories and laptops, at which versions, or which of them are the same server your engineers also connected to a different AI tool.