SUMM

Recent development focus has been on GitHub Copilot, evolving from code completion to chat and now agent mode for deeper, multi-step interactions.
Agent mode allows a user to specify a complete task, interact deeply with the agent (including terminal commands), and can be used to build new apps or perform major refactors.
A comprehensive README serves as the starting point—describing the project, environment setup, database schemas, API endpoints, and workflow diagrams (can be text or images).
With agent mode, Copilot can generate and set up basic working applications in minutes, seeking user permissions when terminal or environment actions are required.

MCP is an open protocol/API enabling LLMs to access external data sources and references, with or without authentication.
MCP allows integration with a wide range of sources, both general and account-specific, and can provide advice from external systems.
MCP servers can be browsed at github.com/modelcontextprotocol/servers; slides and resources are shared at gh.io/fair/mcp.
Integration with MCP is configured in the IDE (e.g., VS Code), so Copilot can leverage different MCPs based on task requirements.

Copilot detects and uses relevant MCPs as per developer instructions; the selection and configuration are performed in VS Code.
MCP servers can be local or remote, providing either local access (e.g., to a database) or secure remote access in protected environments.
Example use case: Mocking data for tests using a Postgres MCP, which is configured and started in VS Code, often with one-click setup.
The developer must fill in connection specifics (such as authentication and host details) in settings.

Enabled MCP servers are visible via a tools icon in VS Code chat, allowing enable/disable and reference management.
Copilot can be directed to perform actions using a specific MCP (e.g., explicitly stating “use the Postgres MCP”).
The workflow involves: getting user permission to connect, connecting in read-only mode, retrieving database schema/data, generating a mock JSON file, and proceeding to build out tests.
Read-only MCPs protect production data integrity by restricting operations to data retrieval only.

Copilot parses prompts, decides whether to use MCPs, asks the user for permission, and interacts with MCP servers to fetch required data.
The process includes retrieving database schemas, table data, parsing results, and incorporating outputs (like files) into the project context.
Iterative workflow is emphasized—initial prompts rarely yield perfect results, and using tools like detailed READMEs and copilot instructions improves effectiveness.
Prompt files and copilot instruction files in VS Code help with reusability and maintain consistent directives.

A GitHub-specific MCP server exists, allowing Copilot to automate tasks (e.g., creating branches, PRs) that would otherwise require manual steps through the CLI or web interface.
Setup involves generating a personal access token, adding it in VS Code, and starting the MCP server with user-specific authentication.
Best practice: Use .github/copilot-instructions.mmd to define standard prompt instructions, coding standards, and to enforce change log creation for transparency in all agent actions.
After code changes, the agent can be instructed to use the GitHub MCP to commit changes to a new branch and open a pull request, automating the end-to-end workflow.

Mention of a new agent (potentially “Gupil” or “Coding Agent”) that allows assigning GitHub issues directly to Copilot, which acts autonomously and provides results in a pull request/session on github.com.
Currently, more unified agent workflows seem suited for enterprise use, but future integration paths are undetermined.
These GitHub Copilot agent features also support MCP configuration through repository settings.
Reference links for further resources and documentation are provided at the end of the session.

Real world MCPs in GitHub Copilot Agent Mode — Jon Peck, Microsoft