Claude Code in VS Code: Extension, Diffs, and Settings

You need VS Code 1.94.0 or later and an Anthropic account: any paid Claude plan (Pro, Max, Team, or Enterprise) or a Claude Console account. Bedrock, Google Cloud, and Foundry users configure the provider through the extension's environment settings instead.

Install from the Extensions view (Cmd+Shift+X / Ctrl+Shift+X, search "Claude Code"). The same extension installs in Cursor and other VS Code forks, or from the Open VSX registry.

The Spark icon marks Claude Code throughout VS Code. Click it in the editor toolbar (top right, visible when a file is open), in the Activity Bar for the sessions list, or ✻ Claude Code in the status bar. The Command Palette also has Claude Code: Open in New Tab. The first time you open the panel, click Sign in and finish in the browser.

One detail causes a lot of confusion: the extension bundles a private copy of the CLI for its panel, but it does not put claude on your shell PATH. To use claude in a terminal, claude mcp add, or claude --resume, install the standalone CLI as well.

The extension sends more than the text you type:

• The open file: Claude sees which file is open in the editor, even with nothing selected. Turn off Attach Open File (v2.1.271 or later) to send only selections.
• Your selection: highlighted code is included automatically, and the prompt footer shows how many lines. Press Option+K (Mac) or Alt+K (Windows/Linux) to insert an explicit reference such as @app.ts#5-10. Click the X on the selection indicator to leave it out.
• @-mentions: type @ and part of a name; matching is fuzzy, so @auth finds auth.js or AuthService.ts. Add a trailing slash for a folder: @src/components/.
• Attachments: paste an image into the prompt box, or hold Shift while dragging files in.

The extension also withholds selected text from some files. If a workspace file matches your files.exclude or search.exclude settings, or is ignored by git (with search.useIgnoreFiles and the extension's respectGitIgnore on, the defaults), Claude receives at most its path.

That filter applies to the chat panel only. When the CLI runs in the integrated terminal, it sends your selection whatever the file. To keep a file such as .env away from Claude in both, add a Read deny rule for its path.

The mode indicator at the bottom of the prompt box shows the permission mode; click it to switch. With Claude Code v2.1.283 or later, Auto is the built-in starting mode, so Claude edits most workspace files and runs routine commands without asking while a classifier blocks risky actions. The other modes are Manual (asks before edits and most shell commands), Edit automatically (edits without asking), and Plan.

In Manual mode, each proposed edit opens as a side-by-side comparison with a permission prompt. You can accept, reject, or tell Claude what to do instead. Since v2.1.275 you can review one change at a time with Accept this change and Reject this change buttons under each hunk, or from the editor's context menu. You can also edit the proposed content directly in the diff before accepting; Claude is told you modified it, so it does not assume the file matches its proposal.

Plan mode makes Claude describe what it will do and wait. VS Code opens the plan as a full Markdown document where you can add inline comments before Claude starts. Type /plan to switch, or /plan fix the auth bug to switch and start planning that task (v2.1.280 or later).

The extension tracks Claude's file edits as checkpoints. Hover over any message to show the rewind button, then choose:

• Fork conversation from here: a new conversation branch, keeping all code changes.
• Rewind code to here: revert file changes to that point, keeping the conversation.
• Fork conversation and rewind code: both.

Checkpoints cover edits made through Claude's file tools, not changes made by shell commands, so commit before large tasks.

Conversations are saved and listed in the sessions view (Spark icon in the Activity Bar), where you can filter and group them. Open several at once as editor tabs (Cmd+Shift+Esc / Ctrl+Shift+Esc for a new tab) or in separate windows. The extension and CLI share conversation history: run claude --resume in a terminal to pick up an extension conversation there. You can also resume cloud sessions from claude.ai in the extension.

The extension reads configuration from two places. Extension settings, in VS Code's settings UI, control the extension's own behaviour. Claude Code settings, in ~/.claude/settings.json and the project's .claude/settings.json, hold permission rules, hooks, MCP servers, and environment variables, and are shared with the CLI and the desktop app.

Put anything you want everywhere (rules, hooks, MCP) in the Claude Code settings files. Use the extension's environmentVariables for things the extension itself must see before it starts, such as ANTHROPIC_BASE_URL and a credential for an LLM gateway, because values only in ~/.claude/settings.json do not reach the extension's own login check.

Add "$schema": "https://json.schemastore.org/claude-code-settings.json" to a settings.json to get autocomplete and validation in VS Code.

Some features exist only in the CLI: the full set of commands and skills, the ! Bash shortcut, and tab completion. To use them without leaving VS Code, open the integrated terminal and run claude (with the standalone CLI installed). The CLI detects VS Code automatically. From an external terminal, run /ide inside Claude Code to connect it.

The connection works through a local MCP server the extension runs, named ide. Through it, the CLI opens diffs in VS Code's native diff viewer, reads your current selection for @-mentions, sends the active file path with each prompt, and can execute cells in a Jupyter notebook. The server binds to 127.0.0.1 on a random port. Each extension activation writes a fresh auth token to ~/.claude/ide/<port>.lock with 0600 permissions, and the CLI must present it to connect.

Only two of its tools are visible to the model: mcp__ide__getDiagnostics, which returns the errors and warnings from the Problems panel, and mcp__ide__executeCode, which runs Python in the active notebook's kernel after confirmation. If your organisation allowlists MCP tools with a PreToolUse hook, include them. In the chat panel, diagnostics come from a separate server, so rules must also name mcp__claude-vscode__getDiagnostics.

Shortcuts depend on focus: when the cursor is in a file the editor is focused, and when it is in the prompt box Claude is. Cmd+Esc (Mac) or Ctrl+Esc (Windows/Linux) toggles between them, and it is the one shortcut to learn first.

• Option+K / Alt+K: insert an @-mention for the current file and selection (editor focused).
• Cmd+Shift+Esc / Ctrl+Shift+Esc: open a new conversation in an editor tab.
• Cmd+Shift+T / Ctrl+Shift+T: reopen the last closed Claude tab.
• Ctrl+Option+F / Ctrl+Alt+F: toggle Focus view, which hides tool calls and thinking behind expandable rows.
• Cmd+N / Ctrl+N: new conversation, only if you enable enableNewConversationShortcut.

Every extension command is in the Command Palette under "Claude Code", including Accept Change at Cursor, Reject Change at Cursor, Show Logs, and Open Walkthrough for a guided tour.

Use a small project with a test command.

1. Install the extension, open the panel from the Spark icon, sign in, and switch the mode indicator to Manual. 2. Select a function, press Option+K / Alt+K, and ask: "add input validation to this function and a test for it". 3. In the diff, accept one change, reject another, and edit one line yourself before accepting the file. 4. Hover over your first message and choose Rewind code to here. Confirm with git status that the edits are gone. 5. Install the standalone CLI, open the integrated terminal, run claude --resume, and pick the same conversation. 6. Add "deny": ["Read(./.env)"] to .claude/settings.json, select text in .env, and confirm Claude cannot read it from either the panel or the CLI.

Pass criterion: per-change review works, the rewind restores the files, the conversation resumes in the CLI, and .env stays unreadable in both.