Copilot Agents, MCP & Agentic Development — assessment

• Whether each developer is on a paid individual plan.
• Whether their IDE supports inline completion.
• Whether the servers are written in a supported language.
• Whether the "MCP servers in Copilot" organisation policy has been enabled — Business and Enterprise must switch it on before MCP can be used.

• A specially scoped token with read-only access to the current repository.
• Your personal access token, with your full permissions.
• An organisation-wide admin token.
• No token; the connection is unauthenticated.

• Only a fixed number of tools may be enabled at once.
• Limiting the toolset reduces attack surface and improves performance at the same time.
• It is required for the server to start.
• Unlisted tools are billed separately.

• It lists vetted servers, which narrows the trust problem without removing it — you still decide what a given server may reach.
• It guarantees every listed server is safe for any use.
• It replaces the need for organisation policy.
• It hosts the servers on GitHub's infrastructure.

Select all that apply.

• You want the result to arrive as a pull request that someone can review normally.
• You want to avoid reviewing the change.
• You want to intervene at each step as the change is made.
• The task is well-specified and you would rather not sit and watch it happen.

• Giving it a narrow, well-defined job and only the tools that job needs.
• Pinning it to the largest available model.
• Giving it access to every tool so it is never blocked.
• Writing the longest possible system prompt.

• Approve it; MCP servers only request what they need.
• Disable MCP entirely; there is no safe configuration.
• Approve it if the server is in the registry.
• Refuse, and grant access only to the specific paths the task genuinely needs.

• It covers those interactions, blocking secrets that appear in AI-generated responses.
• It only applies to commits, never to responses.
• It replaces the need to review what an agent produces.
• It does not apply to AI-generated content.

• How fast it completes tasks.
• What it can reach if it behaves unexpectedly, and whether that damage would be recoverable.
• Which model it is using.
• Whether it usually behaves correctly.

• When the task needs the agent to take actions with tools, rather than just receive a well-structured instruction.
• When you are on a Free plan.
• When the task is shorter.
• When you want it to run automatically on every request.