Basics of GitHub Copilot — Workshop
A practical introduction to GitHub Copilot for anyone looking to start using it for coding.
Pre-Requisites:
GitHub Copilot enabled on your account — check GitHub settings
Latest version of VS Code — required for many newer Copilot features
The GitHub Copilot extension installed in VS Code
A Codebase you want to use to practice talking to Copilot.
Copilot Chat working — open a project, open the chat window, and try: "What is the name of the codebase I have open?"
Duration: ~2 hours
https://docs.google.com/presentation/d/1K-QU4tRBbnq5UH7Z5jlppUS8ALrbZBTXY-RyCYBnqfs/edit?usp=sharing
Welcome to GitHub Copilot Basics! This training takes about two hours to go through, and it's designed so you can either follow along in a live session or work through it on your own at your own pace. There are two hands-on exercises built in so you can practice as you go.
Here's what we're going to cover:
First half: What LLMs and AI agents are, and how GitHub Copilot works — models, modes, tools, and context
Second half: More advanced workflows like running multiple agents and connecting Copilot to external services with MCP
The goal here isn't just to learn Copilot — it's to learn how to think in terms of AI agents. Once you've got that down, you'll be able to pick up any of the other tools out there (Cursor, Claude Code, Windsurf, etc.) pretty easily, because they all work on the same ideas.
Alright, let's get into it.
Workshop Sections
Work through each section in order and complete the exercises as you go!
LLMs vs AI Agents — How LLMs work as next-token prediction machines and how AI agents use them as a brain to take real actions.
Copilot Basics: Models & Modes — How to pick a model, understand premium request costs, and use Ask/Plan/Agent mode to design, plan, and build a game from scratch.
✏️ Exercise 1 - Refactor a File — Practice the full Ask → Plan → Agent workflow on your own codebase by having Copilot refactor a real file.
Multi-Agent Workflows — How to run multiple Copilot chat windows simultaneously and use subagents to parallelize work automatically.
Model Context Protocol (MCP) — How to extend Copilot's toolbox by installing MCP servers that connect it to external services, plus security best practices.
✏️ Exercise 2 - Connect an MCP Server — Install and configure an MCP server of your choice and use one of its tools in a chat.
Wrapping Up
That's a wrap on GitHub Copilot Basics! Here's a quick recap of everything we covered:
What LLMs are and how they work — tokens, vectors, next-token prediction
What AI agents are and how they use LLMs as a brain
The three Copilot modes — Ask, Plan, and Agent — and when to use each
How to choose models and manage your premium request budget
How tools work and how to control which ones Copilot can access
Adding context through ghost text, code highlighting, drag-and-drop, and checkpoints
Running multiple agents simultaneously and using subagents for parallel work
Extending Copilot's capabilities with MCP servers
Everything here applies broadly — the mental model transfers to Cursor, Claude Code, Windsurf, and any other agent you pick up down the road.
Thanks for following along!
Content Index
Core Concepts
Concept | Summary | Section |
|---|---|---|
LLM (Large Language Model) | The core AI technology — a next-token prediction machine trained on massive amounts of text. | |
Token | The unit an LLM actually predicts — smaller than a word, a chunk of text that maps to a number. | |
Vector / Embedding | How text is converted into numbers and placed in a high-dimensional semantic space so the model understands relationships between concepts. | |
Next-token prediction | The fundamental mechanism of LLMs — given some input text, predict the most likely next token. | |
AI Agent | A software layer built on top of an LLM that gives it "hands" — the ability to read/write files, run commands, call APIs, and take real actions. |
Copilot Models & Configuration
Concept | Summary | Section |
|---|---|---|
Model Picker | The dropdown at the bottom of the chat to choose which AI model Copilot uses. Defaults to "Auto". | |
Auto mode (model) | Copilot selects the model for you — comes with a small cost discount. Good default for most tasks. | |
Premium Requests | The usage budget on most Copilot plans. Each model has a multiplier that determines how many premium requests each message consumes. | |
Model Multiplier | A cost signal per model — free, 1x, fractional, or up to 10x — that also roughly indicates capability. | |
Thinking Effort | A low/medium/high toggle on some models that controls how much reasoning the model does before responding. |
Copilot Modes
Concept | Summary | Section |
|---|---|---|
Ask Mode | Conversational-only mode — no file edits or commands. Best for brainstorming, exploring ideas, and building shared understanding before writing code. | |
Plan Mode | Optimized for creating implementation plans. Has its own tool set tuned for research and step-by-step planning. Best used with a powerful model. | |
Agent Mode | Full-power mode — Copilot can edit files, run terminal commands, execute tests, spin up servers, and orchestrate multi-step work. | |
Ask → Plan → Agent workflow | The recommended sequence: design in Ask, plan in Plan, implement in Agent. Produces better results than jumping straight to implementation. |
Tools & Context
Concept | Summary | Section |
|---|---|---|
Tools | Built-in capabilities Copilot can invoke (fetch, file edit, terminal, etc.). Available tools vary by mode. Viewable via the tools icon in the chat. | |
Fetch Tool | Lets Copilot browse the web. Available in Ask mode. Reference it explicitly with | |
| Explicitly invoke a specific tool in the chat by typing | |
Approvals Dropdown | Controls when Copilot asks for confirmation before taking actions: Default (Copilot decides), Bypass (auto-approve), or Autopilot (fully autonomous). | |
Ghost Text | Inline code completions that appear as you type in the editor. Press Tab to accept. | |
Code Highlighting | Select lines in the editor to automatically pin them as context in the chat window. | |
Drag and Drop Files | Drag files or images from the file explorer directly into the chat to add them as context. | |
Context Window Indicator | A UI element (bottom-right of chat) showing how full Copilot's memory is. Start a new chat when it gets too full. | |
Checkpoints | Every edit Copilot makes is tracked as a checkpoint. You can restore any earlier checkpoint to roll back all subsequent file changes. | |
Conversation History | Every message includes the full chat history, so Copilot remembers everything discussed — the more you flesh out in Ask mode, the better it implements later. |
Multi-Agent Workflows
Concept | Summary | Section |
|---|---|---|
Multiple Chat Windows | Open several independent Copilot chat windows side by side — each with its own mode, model, and context — all scoped to the same project. | |
Subagents | A built-in tool ( | |
Isolated Context Window | Each subagent runs in its own context — it only knows what the main agent passes it. Intermediate thinking is discarded after it returns its result. | |
Manual multi-agent | You orchestrate multiple chat windows yourself — good for a small number of independent tasks running in parallel. | |
Automated multi-agent | Ask Copilot to use subagents and it handles all orchestration internally — good for coordinating everything toward a single goal. |
Model Context Protocol (MCP)
Concept | Summary | Section |
|---|---|---|
MCP (Model Context Protocol) | A standard interface for extending Copilot's toolbox by connecting it to external services, APIs, databases, or platforms. | |
MCP Server | A service that exposes tools to Copilot via the MCP standard. Can be installed globally or per-workspace (mcp.json). | |
MCP Marketplace | A built-in browser in Copilot's chat settings for discovering and installing pre-built MCP servers. | |
mcp.json | The workspace-level config file where MCP server configurations are stored when installed per-project. | |
Global vs Workspace Install | MCP servers can be installed globally (available in all projects) or scoped to a single workspace via mcp.json. | |
Deny by Default (MCP security) | Best practice: uncheck all tools when installing a new MCP server, then only enable the specific ones you need — keeps the agent's access surface small. | |
Input Variables | A VS Code feature for referencing secrets (API keys, tokens) in | |
MCP Authentication | MCP servers may require auth tokens or OAuth flows. VS Code provides help for servers in the marketplace. |