Choose and use an agent harness

The Local harness runs interactively in the VS Code extension host and works directly in your active workspace. It can use VS Code built-in tools, extension-provided tools, MCP servers, and the models configured in VS Code, including bring your own key models.

Choose Local for interactive tasks that need immediate feedback or access to editor context, such as diagnostics, test results, terminal output, or selections.

Choose a built-in agent role

Local sessions provide these built-in agent roles:

• Ask: asks questions and provides guidance without making changes to the code.
• Agent: autonomously plans and performs complex coding tasks, edits files, runs commands, and iterates on results.
• Plan: researches a task and creates a structured implementation plan before code changes. Learn more about planning with agents.

You can switch roles during a session from the agent picker.

The Copilot harness is powered by the Copilot SDK and runs locally on your machine in the Agent Host. The Agent Host owns the session independently of the window that displays it, letting you pick up your session from the sessions list in another window or even in the browser.

Setup and authentication

Copilot sessions use the same GitHub authentication context as chat in VS Code. If you use a GitHub Enterprise account for Copilot, the session uses that account. For managed user accounts on GHE.com, complete the setup in Using GitHub Copilot with an account on GHE.com.

Permissions and approvals

The available permission levels depend on the isolation mode:

• Worktree: the permission level is Bypass Approvals and can't be changed.
• Folder: select Default Approvals or Bypass Approvals from the permissions picker. To also use Assisted permissions, turn on chat.assistedPermissions.enabled Open in VS Code Open in VS Code Insiders .

Because Copilot sessions run on the Agent Host, Autopilot is an agent mode rather than a permission level.

Provider-specific capabilities

• Slash commands: enter / in the chat input to view the slash commands available in a Copilot session. For example, use /compact to reduce conversation context or /yolo and /autoApprove to control automatic tool approval.

• Remote control: enter "/remote on" to monitor and steer a running Copilot session from GitHub.com or the GitHub Mobile app. Session history, tool activity, status, approvals, and questions stay synchronized. Remote control requires GitHub authentication and a workspace that maps to a GitHub repository. Enter "/remote off" to stop sharing the session.

• Research agent (Preview): in VS Code Insiders, enter /research <topic> to produce a detailed Markdown report with citations from your codebase, relevant GitHub repositories, and the web. For research that feeds into an implementation plan, use the Plan agent. For focused research that returns results to the current conversation, use subagents.

• Terminal integration: open the GitHub Copilot CLI terminal profile, run Chat: New Copilot CLI Session from the Command Palette (⇧⌘P (Windows, Linux Ctrl+Shift+P)), or enter copilot in an integrated terminal. VS Code adds the session to the sessions list. To continue an existing Copilot session from the terminal, right-click it and select Resume in Terminal.

Limitations

Copilot sessions don't have access to every VS Code built-in or extension-provided tool. Enabled client-side tools are available to the agent only while VS Code is connected to the session, and you manage which tools are enabled from Manage tools for the Copilot harness. Copilot sessions can currently access only local MCP servers that don't require authentication.

Claude sessions use Anthropic's Claude Agent SDK and can run autonomously on your workspace. VS Code integrates the harness through its SDK while keeping session management, chat, and code review in VS Code.

Setup and authentication

Claude support is enabled by default. Turn it on or off with github.copilot.chat.claudeAgent.enabled Open in VS Code Open in VS Code Insiders .

Claude supports two authentication and billing options:

• GitHub Copilot subscription: sign in to GitHub to use Copilot-routed models. Usage is billed through your Copilot subscription.
• Anthropic credentials: use an Anthropic API key or Claude Code OAuth token. Usage is billed by Anthropic.

When both authentication methods are available, the model picker groups models by Anthropic and Copilot. The model you select determines the provider and billing method for the next turn. You can switch between Anthropic-native and Copilot-routed models in an existing Claude session.

To use Claude without signing in to GitHub (Experimental), set ANTHROPIC_API_KEY in your environment or in the env object in ~/.claude/settings.json. Alternatively, set CLAUDE_CODE_OAUTH_TOKEN to a token created with claude setup-token. Learn more about Claude Code authentication.

Enable chat.agentHost.allowSignedOutWhenUsable Open in VS Code Open in VS Code Insiders to open the Agents window while signed out of GitHub. The model picker only shows Anthropic-native models until you sign in. After you sign in to GitHub, Copilot-routed models are also available.

Permissions and approvals

Claude supports these permission modes:

• Edit automatically: apply changes without asking for approval.
• Request approval: ask before applying changes.
• Plan: outline the approach before implementation.

Provider-specific capabilities

Enter / in the chat input to view commands for managing Claude-native agents, hooks, memory files, and code review. Learn more about Claude subagents and Claude hooks.

The Codex harness uses OpenAI Codex for interactive and background coding tasks. It runs through the OpenAI Codex extension or, experimentally, on the Agent Host. VS Code provides session management, chat, and code review for both integrations.

Setup and authentication

Codex is not listed by default. Complete one of these options before you select it. You don't need both:

• Use the OpenAI Codex extension in the Chat view: install and enable the OpenAI Codex extension.
• Use Codex on Agent Host in the Agents window (Experimental): enable chat.agentHost.codexAgent.enabled Open in VS Code Open in VS Code Insiders .

To use the Agent Host implementation in the Chat view, also enable chat.editor.codex.preferAgentHost Open in VS Code Open in VS Code Insiders .

On the Agent Host, Codex supports two authentication and subscription options:

• GitHub Copilot subscription: sign in to GitHub to use Copilot-backed models. This option requires Copilot Pro+.
• ChatGPT subscription: open the account menu and select Sign in to ChatGPT. A free ChatGPT account is sufficient.

When both accounts are signed in, the model picker groups models by Copilot and ChatGPT. Your selection determines which subscription is used, and VS Code saves that provider with the session.

To use Codex without signing in to GitHub (Experimental), sign in to ChatGPT and enable chat.agentHost.allowSignedOutWhenUsable Open in VS Code Open in VS Code Insiders . The desktop Agents window then shows ChatGPT-backed models while signed out. Copilot-backed models prompt you to sign in to GitHub, and the browser-based Agents window still requires GitHub sign-in.

Permissions and approvals

On the Agent Host, Codex provides these approval presets:

• Default Permissions: read and edit workspace files and run routine local commands. Codex asks before using the internet or accessing resources outside the workspace.
• Auto-Review: use the same workspace access as Default Permissions, but send approval requests to an automatic reviewer instead of prompting you.
• Full Access: edit files outside the workspace and use the internet without asking.

The Cloud harness runs on remote infrastructure and works with a GitHub repository. The agent implements the task on a branch and opens a pull request for review. Choose Cloud for well-scoped tasks that can run without access to your local editor context, terminal output, or extension-provided tools.

VS Code supports:

• Copilot cloud agent for implementing features, addressing review feedback, and creating pull requests.
• Claude and Codex cloud agents for provider-specific capabilities. Third-party cloud agents are currently in preview.

To use Claude or Codex in the cloud, turn on support in your Copilot account settings. See Managing policies for third-party cloud agents. You don't need the provider's VS Code extension for a cloud session.

Start a cloud session

1. Open the Chat view and select New Chat.

2. Select Cloud from the Session Target control.

3. Choose the cloud provider and, when available, a custom agent and model.

4. Enter a prompt and submit it.

The session runs remotely and appears in the sessions list. Sessions that you create by assigning an issue or pull request to a cloud agent on GitHub.com also appear in VS Code.

You can also select a GitHub repository when you start a session in the Agents window, or hand off an existing session to the Cloud harness. In a Copilot session, enter /delegate to continue the task in the cloud.

Cloud sessions use the tools, MCP servers, and models configured by the cloud service. They can't access VS Code built-in tools or local runtime context.