Claude Code in VS Code vs Terminal: Features and Settings

Of the two main ways to run Claude Code inside VS Code, only the extension’s chat panel withholds text you select in a gitignored file, such as a .env file holding API keys: by default Claude gets that file’s path at most. Only the CLI takes a shell command typed after !, such as ! npm test, runs it without Claude interpreting or approving it, and adds the output to the conversation; the chat panel has no ! shortcut. Differences like these two decide most of the Claude Code in VS Code vs terminal choice. Both are documented in Anthropic’s VS Code guide and its interactive mode reference as of September 2026.

The extension’s panel runs its own bundled copy of the CLI, so the two surfaces differ mainly in the interface, and the two copies can run different versions. I build and maintain this site with the Claude Code extension in VS Code on Windows 11, and from that setup the surface is the smaller decision and the settings file the bigger one.

Both surfaces read the same settings files, so switching surface cannot rescue a setting that sits in the wrong file. My own user settings file lists two MCP servers, and Claude Code reads neither from it: the Ahrefs tools in my sessions come from a separate entry in ~/.claude.json.

Claude Code in VS Code vs Terminal: Which Should You Use?

Use the VS Code extension to accept or reject Claude’s edits one change at a time in Manual mode, comment on plans inline and keep conversations open in tabs; run claude in the integrated terminal for the full command set, the ! shortcut, tab completion and /config. Both read ~/.claude/settings.json and share conversation history.

Checked on 30 September 2026 against the extension vs CLI section of the VS Code guide and the pages it links, for Claude Code v2.1.283 or later:

Capability	VS Code extension (chat panel)	CLI in the integrated terminal
Built-in commands and skills	A subset; type / to see yours	All of them
Shell commands from the prompt	No !; @terminal:name pulls in a terminal’s output	! runs a command and adds its output
Tab completion	No	Yes
/config settings menu	Not opened; use the extension’s settings or a settings file	Yes
Reviewing an edit in Manual mode	A diff with per-change Accept and Reject buttons (v2.1.275+)	VS Code’s diff viewer while connected; diffTool can keep it in the terminal
Plans	Mode indicator or /plan (v2.1.280+); comment inline on the Markdown plan	Shift+Tab or /plan
Rewind (checkpoints)	Hover a message: fork, rewind code, or both	/rewind, or Esc twice on an empty prompt
Selection in a gitignored or excluded file	Withheld by default; Claude gets the path at most	Sent whatever the file; a Read deny rule blocks it
Pasting a long brief	Pasted text stays visible in the prompt box	A large paste collapses to a placeholder
Parallel conversations	Tabs, windows and a sessions list with groups and filters	claude in a second terminal tab, or agent view (claude agents, research preview)
Saving a “don’t ask again” approval	You pick the file, including the shared project file	Your local settings file only
Remote Control	/remote-control or /rc; no name argument or QR code	Also claude remote-control server mode, --name and a QR code
MCP servers	/mcp dialog; adds and removes servers from v2.1.261	claude mcp add, or /mcp
Where it comes from	Bundled in the extension; claude is not added to your PATH	The standalone install

The extension also has an in-between option: turn on Use Terminal (claudeCode.useTerminal) or run Claude Code: Open in Terminal from the Command Palette, and it opens the CLI-style interface in place of the chat panel. Anthropic’s guide does not say which of the CLI-only rows above that mode changes.

Which permission mode each surface starts in, and how to cut repeat prompts, is covered in why Claude Code keeps asking for permission. The commands from the CLI’s full set that a marketer is most likely to reach for are in my Claude Code commands cheat sheet, which links Anthropic’s complete table.

Two Settings Layers: VS Code’s and Claude Code’s

Anthropic’s VS Code guide splits configuration in two (configure settings):

• “Extension settings in VS Code: control the extension’s behavior within VS Code”
• “Claude Code settings in ~/.claude/settings.json: shared between the extension and CLI”

Anthropic’s settings docs add that the terminal, the VS Code and JetBrains extensions and the desktop app “all read the same settings files”. The diagram maps which surface reads which layer, along with the one-session launch flags that exist only when you start claude yourself.

Two user-level files share the name settings.json. VS Code keeps its own user settings, the extension’s included, in %APPDATA%\Code\User\settings.json on Windows (VS Code settings docs). Claude Code’s user settings file is ~/.claude/settings.json, which on Windows means %USERPROFILE%\.claude\settings.json.

Layer 1: Claude Code VS Code Extension Settings

Open them with Ctrl+, (Cmd+, on Mac), then Extensions, then Claude Code. Each ID starts with claudeCode., and the defaults come from the extension settings table:

Setting	Default	What it controls
preferredLocation	panel	Where Claude opens: sidebar (right) or panel (a new tab)
useTerminal	false	Opens the CLI-style interface (terminal mode) instead of the graphical panel
autosave	true	Saves files before Claude reads or writes them
attachOpenFile	true	Adds the file open in the editor to your messages (v2.1.271 or later)
respectGitIgnore	true	Leaves .gitignore patterns out of file searches and selection context
environmentVariables	[]	Environment variables for the Claude process the extension starts
claudeProcessWrapper	unset	The executable that launches the Claude process, such as a separately installed claude binary
initialPermissionMode and allowDangerouslySkipPermissions	unset and false	The starting permission mode, and whether Bypass permissions appears in the mode selector, which the docs say to use only in sandboxes with no internet access

The docs steer shared configuration out of this layer: next to environmentVariables they say “Use Claude Code settings instead for shared config.”

Layer 2: Claude Code Settings Files Both Surfaces Read

File	Applies to	What goes in it	Who writes it
~/.claude/settings.json	You, in every project	Default model, effort per model, permission rules, env, hooks, Remote Control at startup	You, plus /model, /effort and the extension’s model picker
.claude/settings.json	Everyone in the project, once committed	Team permissions, hooks, plugins	You or your team
.claude/settings.local.json	You, in this project	Personal overrides and “Yes, and don’t ask again” approvals	You and Claude Code
~/.claude.json	You	Sign-in, MCP servers at user and local scope, trust decisions, global keys such as autoConnectIde and diffTool	Claude Code itself, through /config, claude mcp add and the /mcp dialog; the docs say you don’t need to edit it
.mcp.json at the project root	Everyone in the project, once committed	Project-scoped MCP servers	claude mcp add --scope project, or you

Sources: settings files and who they affect, MCP installation scopes and the global config keys, which Claude Code ignores anywhere except ~/.claude.json.

When a key appears in more than one file, Claude Code uses the value from the highest level that sets it: managed settings, then command-line flags, then .claude/settings.local.json, then .claude/settings.json, then your user settings file. Lists merge across files instead of overriding, except four model keys with their own rules: fallbackModel, modelPicker, an availableModels list set in managed settings, and modelSettings (settings precedence).

Adding "$schema": "https://json.schemastore.org/claude-code-settings.json" to the top of ~/.claude/settings.json gives you autocomplete and inline validation in VS Code. The docs warn the schema “can lag behind the newest CLI releases”, so a warning on a newly documented key is not proof of an error.

My Settings Map as of 30 September 2026

This is layer 2 on my machine, with Claude Code v2.1.283 or later. The rest of my setup, from slash commands to auto memory, is in the Claude Code setup behind this site.

File	Key	My value	What it does
~/.claude/settings.json	model	"opus"	New sessions start on the latest Opus, which the model aliases table maps to Opus 5.5 on the Anthropic API
~/.claude/settings.json	modelSettings	xhigh for claude-opus-5, medium for claude-opus-5-5	One saved effort level per model
~/.claude/settings.json	remoteControlAtStartup	true	Every new interactive session connects to Remote Control on its own
~/.claude/settings.json	agentPushNotifEnabled	true	Lets Claude send a push notification to a phone signed in to the Claude app; the settings reference says pushes arrive while Remote Control is connected
~/.claude/settings.json	mcpServers	Figma and Ahrefs entries	Nothing: settings.json does not read this key
~/.claude.json	top-level mcpServers	Ahrefs, "type": "http"	A user-scope server, and the source of the Ahrefs tools in my sessions
~/.claude.json, this repo’s entry	mcpServers	None	No local-scope servers for this project
.mcp.json	None	No file	No project-scope servers
~/.claude/settings.json, .claude/settings.json and .claude/settings.local.json	permissions	Allow rules in all three files	Pre-approved commands, fetch domains and MCP tools, counted per file in the permissions guide linked above
Any settings file	hooks	None	I do not use Claude Code hooks

My sessions run with Git Bash and PowerShell as shells; Windows install routes and shells are compared in Claude Code on WSL vs native Windows.

Where Claude Code Reads MCP Servers From

Anthropic’s MCP docs list three scopes for a server you add yourself (scope hierarchy and precedence):

Scope	Stored in	Loads in
Local, the default for claude mcp add	~/.claude.json, under the project’s path	That project only
Project	.mcp.json at the project root	That project, for everyone who has the file, after a one-time approval
User	~/.claude.json, at the top level	Every project on your machine

When one server name appears in more than one scope, local beats project and project beats user. Plugin-provided servers and claude.ai connectors rank below all three.

~/.claude/settings.json is missing from that table. The troubleshooting table on Anthropic’s debug your configuration page lists the symptom “MCP servers added under mcpServers in settings.json never appear” with the cause “settings.json does not read an mcpServers key”. The VS Code guide’s settings section points the other way, naming “allowed commands, environment variables, hooks, and MCP servers” as what that file is for. The MCP keys your user, project or local settings file does accept approve or restrict servers, such as enableAllProjectMcpServers, allowedMcpServers and deniedMcpServers.

My user settings file has exactly this problem: its mcpServers block holds Figma and Ahrefs entries, both written as launch commands, and no Figma tools load in my sessions.

To move a server out of the settings file:

1. See what loaded: /mcp in the chat panel, or claude mcp list in the terminal, which shows each server’s health status.
2. With the standalone CLI installed, re-add each entry with claude mcp add-json <name> '<json>' --scope user, passing the object inside mcpServers rather than the wrapper (add servers from JSON). An entry with a url and no type needs a type such as "http" added first, since Claude Code reads an entry with no type as a stdio server. The same page’s tip reads “Make sure the JSON is properly escaped in your shell”; its examples wrap the JSON in single quotes, which Git Bash passes through as written. On Windows, VS Code’s integrated terminal opens PowerShell by default (terminal profiles); if the command rejects the JSON there, switch the profile to Git Bash with Terminal: Select Default Profile, or add the server from the /mcp dialog in step 3.
3. Or add it from the chat panel’s /mcp dialog, which from v2.1.261 saves to the same configuration as claude mcp add (MCP in the VS Code guide). A guide that says servers can be added only from the terminal is out of date for v2.1.261 and later.
4. Delete the old block from settings.json, so the file stops listing servers that never load.
5. Start a new conversation and check that the server shows Connected in /mcp. The VS Code guide says changes from the dialog or the terminal “take effect in conversations you start afterwards.”

How to Change the Model in the Claude Code VS Code Extension

In the chat panel, click the model name at the bottom of the prompt box, or open the / command menu and choose Switch model. When the model supports effort levels, the picker also shows an Effort row. Per the guide’s prompt box section, “When you pick a level other than max, Claude Code saves it for the current model as your default, under modelSettings in your user settings; max applies to the current session only.” The model name button and the Effort row need v2.1.257 or later.

In the terminal, the same choices look like this, per the model configuration docs:

What you want	Command	Saved for new sessions?
Switch model and make it your default	/model opus, or Enter in the /model picker	Yes, as model in ~/.claude/settings.json
Switch model for this session only	s on a row of the /model picker	No
Set effort	/effort high, or the slider in /model	Yes, per model in modelSettings, except max
One launch only	claude --model sonnet --effort low	No

Here are the lines of my ~/.claude/settings.json that set the model, effort and Remote Control for both surfaces, with the other keys left out:

{
  "model": "opus",
  "modelSettings": {
    "claude-opus-5": { "effortLevel": "xhigh" },
    "claude-opus-5-5": { "effortLevel": "medium" }
  },
  "remoteControlAtStartup": true
}

What those lines do, according to the effort level docs and the modelSettings entry:

• The panel and the terminal both start new sessions on the model opus resolves to, shown in the map above.
• Claude Code resolves modelSettings one model at a time. A new session on Opus 5.5 starts at medium, the same level Opus 5.5 uses when nothing is saved for it, and the xhigh entry applies only in a session running Opus 5.
• Saving per model keeps an old level from following you to a new model, which matches the docs’ advice: “When you move from Opus 5 to Opus 5.5, start at medium rather than carrying over the level you used on Opus 5.” In Anthropic’s testing, “Opus 5.5 at medium matches or exceeds Opus 5 at high on coding and knowledge-work evaluations.”
• A top-level effortLevel in your user settings file is the older format, and the docs say it “doesn’t count for Opus 5.5”.
• A conversation you reopen with claude --resume keeps the model it was using, per the sessions docs, so it does not pick up a default you changed since.

Which model to run for which job, and what each costs per job, is in Claude Sonnet vs Opus vs Haiku.

How to Run Claude Code in VS Code’s Integrated Terminal

The chat panel needs only the extension: install Claude Code from the Extensions view (Ctrl+Shift+X, or Cmd+Shift+X on Mac), open a file and click the Spark icon in the editor toolbar, per the install section of Anthropic’s VS Code guide. The terminal route takes more steps:

1. Install the standalone CLI from Anthropic’s setup page. The extension bundles a private copy of the CLI for its chat panel, and installing it “does not put claude on your shell PATH.”
2. Open the integrated terminal with Ctrl+` (Cmd+` on Mac) and run claude. Per the guide’s run CLI in VS Code section, the CLI “automatically integrates with your IDE for features like diff viewing and diagnostic sharing.” It connects through a local MCP server named ide that the extension runs and /mcp hides.
3. Run /terminal-setup once so that Shift+Enter adds a new line instead of sending. The terminal configuration docs say that in VS Code it also turns off the integrated terminal’s GPU acceleration to prevent garbled text.
4. From a terminal outside VS Code, run /ide to connect to the editor, or turn on Auto-connect to IDE (external terminal) in /config, which saves autoConnectIde to ~/.claude.json.
5. Proposed edits open in VS Code’s diff viewer while the CLI is connected, because diffTool defaults to "auto". To keep them in the terminal, set Diff tool to terminal in /config; that row appears only while the CLI is connected to VS Code.
6. To carry on a panel conversation, run claude --resume and pick it from the list. The same guide says “The extension and CLI share the same conversation history.”
7. Check both version numbers, because they come from separate copies: claude --version in the terminal, and /status in the panel for the build a conversation runs.

On 30 September 2026, claude --version on my machine printed 2.1.283, while the newest extension build installed was 2.1.285. A feature with a version floor can reach one surface first: Sonnet 5.5 needs Claude Code v2.1.284 or later, so a 2.1.283 CLI cannot run it.

Running claude inside a VS Code terminal also reinstalls the extension if you removed it. To stop that, turn off Auto-install IDE extension in /config (the autoInstallIdeExtension key).

Claude Code in VS Code vs Terminal: Choose the Surface, Then the File

The surface is the choice you can change mid-task, since claude --resume carries a panel conversation into the terminal. The file decides whether a setting applies at all, as the mcpServers block in my user settings file shows.

Open ~/.claude/settings.json and search it for mcpServers: every server listed there has to move to ~/.claude.json or .mcp.json, using the steps above, before it loads.