Claude Code in VS Code: extension, terminal, or both

What the extension actually is

A graphical Claude Code panel that lives in the editor. The Spark icon marks it: in the editor toolbar when a file is open, in the Activity Bar for the sessions list, and in the status bar as Claude Code when nothing is open. Drag the panel into the secondary sidebar to keep it beside your code, into the primary sidebar, or into the editor area as an ordinary tab.

• Plan review as a document. In plan mode the plan opens as a full Markdown file you can annotate inline before Claude starts. Commenting on a plan in an editor is a different activity from arguing with a paragraph in a terminal.
• Diffs before the edit. Every proposed change opens side by side with an accept, reject or redirect prompt. Edit the proposed content in the diff view and Claude is told you modified it, so it does not assume the file matches its own proposal.
• Selection is context automatically. Highlighted code is visible to Claude without pasting; the footer shows how many lines. The eye-slash icon hides the selection when you would rather it did not.
• Checkpoints. Hover any message to fork the conversation, rewind the code to that point, or both.
• Sessions in tabs and windows. Open in New Tab or Open in New Window runs several conversations with independent context. A blue dot means a permission request is waiting; orange means Claude finished while the tab was hidden.

Setting it up

curl -fsSL https://claude.ai/install.sh | bash
exec $SHELL -l
claude --version

Terminal versus extension, row by row

The two surfaces share conversation history, so this is not a fork in the road. Start in the panel, then run claude --resume in the terminal to pick the same conversation up in an interactive picker.

The shortcuts worth learning

A handful of extension settings change the feel more than the shortcuts do. useTerminal swaps the panel for the CLI interface. initialPermissionMode sets the mode every new conversation starts in, including plan. useCtrlEnterToSend stops a stray <kbd>Enter</kbd> firing a half-written prompt. respectGitIgnore is on by default and keeps ignored paths out of file search.

The IDE bridge, and what it sends

When the extension is active it runs a small local MCP server named ide, and the CLI connects to it automatically. This is the machinery behind native diffs, selection context, and notebook cell execution. It is worth understanding because it changes what leaves your editor.

• Selection and open file travel with each prompt. The transcript shows a Selected N lines from <file> line when it happens.
• Two tools reach the model: mcp__ide__getDiagnostics, which returns the Problems panel contents, and mcp__ide__executeCode, which runs Python in the active Jupyter kernel. The rest of the server is internal RPC and is filtered out before the tool list reaches Claude.
• Notebook execution always asks. The code is inserted as a new cell and a native Quick Pick asks you to execute or cancel. Dismissing it returns an error to Claude and nothing runs.
• The socket is loopback only, on a random port, with a fresh auth token per activation written to ~/.claude/ide/<port>.lock at mode 0600.

When the terminal still wins

• Several agents at once. Terminal tabs, tmux panes and worktrees handle this naturally. The extension can open sessions in tabs, but an editor window is a worse dashboard than a terminal multiplexer once you pass three.
• Long autonomous runs. An agent in a window you are not looking at is a feature. An IDE panel invites you to watch it, which is the most expensive way to use one.
• Remote and headless. SSH into a box and run it. No extension host, no port forwarding, no Remote-SSH round trip.
• Scripting. claude -p "..." --output-format json in a pipeline has no IDE equivalent.
• The muscle-memory bits. The ! bash shortcut and tab completion are terminal-only, and if you rely on them the panel feels slower even when it is not.

# one isolated worktree and branch per session
claude --worktree feature-auth
claude --worktree feature-search

Questions people ask

Keep reading