In a nutshell
Copilot Chat is the synchronous conversation surface for Copilot.
Here it refers to the experience of asking, consulting, and requesting Copilot in the IDE, GitHub.com, and Mobile.
Cloud Agent, which fully delegates implementation, and Copilot CLI, which runs in the terminal, are separate topics. Chat is first and foremost a “UI for conversing on the spot.”
Three Chats
Even when called “Copilot Chat,” the experience varies a bit depending on where you open it.
| Chat | Where | Role |
|---|---|---|
| IDE Chat | VS Code, Visual Studio, JetBrains, etc. | Converse with Ask / Plan / Agent while using the local workspace and tools |
| GitHub.com Chat | GitHub in the browser | Ask questions and make requests from issue, PR, and repo context. Also serves as an entry point for Cloud Agent |
| Mobile Chat | GitHub Mobile | Review and instruct on the go. Also an entry point for launching agent sessions from issues / PRs |
GitHub.com / Mobile are convenient entry points. However, the same modes, built-in agents, skills, and tools as the IDE are not always available.
Three IDE Modes
VS Code’s Copilot Chat has built-in agents for different purposes.
| Mode | Underlying mechanism | What does it do? |
|---|---|---|
| Ask | Ask built-in agent | The first mode to appear as GitHub Copilot. Provides answers to questions. Recommended for new team members getting up to speed with the codebase |
| Plan | Plan built-in agent | Create implementation-ready documentation through interactive conversation with GitHub Copilot. Use Agent mode or Coding Agent to implement based on the created documentation |
| Agent | Raw agent | Acts as a semi-autonomous pair programmer: internally analyzes requests, proposes implementation, and validates proposals. Can interact with systems outside GitHub using MCP servers |
Even within the same Chat UI, what is “allowed” differs between Ask / Plan / Agent.
IDE Chat Operations
VS Code’s Copilot Chat has built-in skills, prompts, instructions, hooks, MCP servers, and other components visible from the start.
| Operation | Where to look | What you can do |
|---|---|---|
| Check components | Cog icon at top-right of Chat → Agent Customizations | View Agents, Skills, Instructions, Prompts, Hooks, MCP Servers, and Plugins |
| Invoke agent / prompt | / in the Chat input | Launch Ask / Plan / Agent, built-in commands, and prompts |
| Add context | # or Add Context | Pass #file, #codebase, folder, symbol, terminal output, web fetch, and more |
| Check tools | Configure Tools button | Choose built-in tools, MCP tools, and extension tools available to the Agent |
Instructions / Skills / MCP / verification policies created in the PLAN phase are passed here as context and tools to Chat / Agent.
Models and hosting
In Chat / Agent you can pick from several models. The key point: the provider that built the model and where it actually runs (the host) are not the same thing. GitHub has agreements with each host so your data is not used for training.
| Model family | Provider (built it) | Where it actually runs (host) |
|---|---|---|
| GPT-5.5 / 5.4 / 5.3-Codex / mini / nano | OpenAI | OpenAI and GitHub’s Azure |
| Raptor mini (GPT-5 mini fine-tuned by Microsoft) | OpenAI × Microsoft | GitHub-managed Azure OpenAI tenant |
| Claude Opus / Sonnet / Haiku | Anthropic | AWS Bedrock / Anthropic PBC / Google Cloud |
| Gemini 3.1 Pro / 3.5 Flash / 2.5 Pro | Google Cloud Platform (GCP) | |
| MAI-Code-1-Flash | Microsoft (first-party) | Azure (GitHub’s tenant) |