MCP VS Code
Lets an AI agent such as Claude Code work inside your VS Code window, the way you do: run tasks, terminal
commands and Command Palette commands, navigate and refactor code with the language servers, debug, search
files, read the Problems and Output panels, and manage tabs, tests and extensions. Everything it does is
visible in the window. The extension serves the
Model Context Protocol on 127.0.0.1:47391.
Features
- Tasks: run any task from
tasks.json or auto-detected npm scripts, wait for exit or for a
"ready" line, read the output, stop it.
- Terminals: run shell commands in named terminals and read their exit code and output, including
commands you type yourself.
- Commands: run any Command Palette command by id, such as Format Document or Reload Window.
- Code intelligence: go to definition, find references, hover, symbols, call hierarchy, safe rename and
quick fixes, all from the language servers.
- Debugging: start a launch configuration, set breakpoints and logpoints, step, read the call stack and
variables, evaluate expressions.
- Search: search the workspace text like the Search view (Ctrl+Shift+F), with regex, case, whole-word
and include/exclude globs.
- Problems and Output: read errors and warnings, and the Output channels of extensions (Git,
TypeScript, ESLint, ...).
- Editors: list, open and close tabs; format and save files; read the cursor and the selected text.
- Tests: discover and run tests in the Test Explorer.
- Extensions: list, install and uninstall extensions.
- Also listed under
@mcp in the Extensions view, so VS Code chat agents can use the same tools.
Settings and commands
| Setting |
Default |
Does |
mcpVscode.enabled |
true |
Serve the MCP endpoint from this window |
mcpVscode.port |
47391 |
Port of the endpoint; the first window to bind it serves |
mcpVscode.disabledTools |
[] |
Tools the endpoint neither lists nor runs |
mcpVscode.disabledTools restricts what agents can do, for example:
"mcpVscode.disabledTools": ["run_command", "run_in_terminal", "install_extension", "uninstall_extension"]
- User and workspace values add up. A workspace can disable more tools, but its settings cannot re-enable a
tool disabled in user settings.
- Changes apply at once, without a restart. A refused call tells the agent the tool is disabled. Clients
that cached the tool list see the shorter list after reconnecting (
/mcp in Claude Code).
run_command and run_in_terminal can reach most of what the other tools do, such as installing an
extension through its command. Disable them too when the goal is a real lockdown.
Commands: MCP VS Code: Copy Claude Code Setup Command (copies the claude mcp add ... line),
Restart Server, Show Log.
Tests
npm run test:live starts an isolated VS Code window with this extension loaded from source and calls
every tool over MCP. See tests/live/README.md.
Limits
- One window serves the port; other windows show a disconnected status bar item (click to retry).
- Claude Code connects at startup; after starting VS Code later, reconnect with
/mcp.
- Not readable through the VS Code API: notifications, test results, and the output of tasks VS Code must
run itself (
dependsOn, ${input:...}, inline problem matchers).
- Output channels are read from VS Code's log files, which is not a public API (
src/unofficial). The
layout is verified before use; when it no longer matches, the tools say "API is no longer supported".
Core channels such as Tasks or Extension Host are not covered.
- Language tools answer only as well as the installed language extension. Right after a file is opened its
server may still be loading, and the tools say so.
- Terminal output needs shell integration (PowerShell, bash, zsh, fish), not cmd.exe. PowerShell reports
every failed command as exit code 1.
- Only local requests are served: non-local
Host headers and browser Origin headers are refused.
Full command list
Paths may be absolute or workspace-relative; in a multi-root workspace a relative path may start with a
folder name. Output tools page with since / max_lines and filter with a case-insensitive regex
filter; their reply ends with the Next since to pass on. wait_for_exit, wait_for_pattern and
timeout_seconds make a call block. Position tools take path and a 1-based line, plus column or the
symbol name on that line. Editing tools save the changed files unless save: false.
| Tool |
Does |
list_tasks |
Tasks VS Code can run, with the state of tracked runs |
run_task |
Start a task; optionally wait for exit or an output pattern; restart stops a running one first |
task_status |
State, exit code and output of a run; without name, all tracked runs |
close_task |
Stop a task's process tree and close its terminal |
run_in_terminal |
Run a shell command in a named terminal (default MCP) and cwd; returns exit code and output |
read_terminal |
Output of a terminal's latest (or a matching) command; without terminal, all terminals |
clear_terminal |
Clear a terminal and its captured commands |
list_commands |
Command ids, with titles of extension-contributed commands; filter narrows |
run_command |
Run a command by id with JSON args; file opens and focuses a file first (for editor commands) |
search_in_files |
Matching lines grouped by file; is_regex, case_sensitive, whole_word, include, exclude, show |
find_definition |
Definition, type definition, implementation or declaration (kind) of the symbol at a position |
find_references |
Every reference to the symbol at a position, with the line text |
get_hover |
Type, signature and documentation of the symbol at a position |
document_symbols / workspace_symbols |
Outline of a file; symbols by name across the workspace |
call_hierarchy |
Callers (incoming) or callees (outgoing) of the function at a position |
rename_symbol |
Rename the symbol at a position in every file that uses it |
code_actions |
Quick fixes and refactorings for lines (kind narrows); apply: <number> applies one |
format_document / save_files |
Format a file with its configured formatter; save one or all unsaved files |
get_selection |
File, cursor and selected text of the active editor |
get_diagnostics |
Problems panel entries, errors first; filter by path and severity |
list_output_channels / read_output_channel |
Output channels of extensions; last lines of one, with filter |
list_debug_configs / start_debugging |
Launch configurations; start one, optionally wait_for_pause |
breakpoints |
list, add (with condition or log_message) or remove breakpoints |
debug_control |
continue, pause, step_over, step_into, step_out, restart; waits for the next pause |
debug_inspect / stop_debugging |
Call stack, variables and an evaluated expression; stop all sessions |
get_workspace |
Open folders, workspace file, trust and remote state |
list_tabs |
Tabs per editor group, unsaved and preview flags, active cursor and selection |
open_file / close_tab |
Open a file at a line; close a file's tabs (save: true for unsaved ones) |
test_explorer |
Test controllers with tests, or the tests of one file |
refresh_tests / run_tests |
Rediscover tests; run all, failed, last or current_file and wait |
cancel_test_run / clear_test_results |
Cancel the test run; clear results |
list_extensions |
Installed extensions: active, inactive, disabled or uninstalled (until reload) |
install_extension / uninstall_extension |
Install by id or .vsix path; uninstall by id |
Useful run_command ids: workbench.action.reloadWindow (the connection drops until VS Code is back),
editor.action.organizeImports (with file), workbench.action.closeAllEditors. Arguments are plain JSON,
so commands that need Uri or Position objects are covered by the dedicated tools above instead.
License
MIT. See the LICENSE file.
| |
