BashTerm

BashTerm

Enables executing commands in visible VSCode terminal tabs with full output capture, supporting long-running processes, interactive input, and isolated sessions for parallel agents.

Category
访问服务器

README

vscode-terminal-mcp

npm version

MCP server that executes commands in visible VSCode terminal tabs with full output capture. Unlike inline execution, every command runs in a real terminal you can see, scroll, and interact with.

Key Features

  • Visible Terminals: Commands run in real VSCode terminal tabs, not hidden processes. You see everything in real time.
  • Session Reuse: The run tool automatically reuses idle sessions, creating new terminals only when needed.
  • Long-Running Support: Fire-and-forget execution with waitForCompletion: false, then poll output incrementally with read.
  • Subagent Isolation: Tag sessions with agentId to keep parallel agent workloads separated.

Requirements

  • VS Code 1.93+ (for Shell Integration API)
  • Node.js 20+

Getting Started

Claude Code

claude mcp add BashTerm -- npx vscode-terminal-mcp@latest

VS Code / Copilot

Add to your .vscode/mcp.json:

{
  "servers": {
    "BashTerm": {
      "type": "stdio",
      "command": "npx",
      "args": ["vscode-terminal-mcp@latest"]
    }
  }
}

<details> <summary>Cursor</summary>

Add to your .cursor/mcp.json:

{
  "mcpServers": {
    "BashTerm": {
      "command": "npx",
      "args": ["-y", "vscode-terminal-mcp@latest"]
    }
  }
}

</details>

<details> <summary>Claude Desktop</summary>

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "BashTerm": {
      "command": "npx",
      "args": ["-y", "vscode-terminal-mcp@latest"]
    }
  }
}

</details>

Your First Prompt

After installation, try asking:

Run ls -la in the terminal

You should see a new terminal tab open in VSCode with the command output.

Screenshots

Running a command with run

Run command output

Permission dialog for exec

Exec permission dialog

Exec result with clean output

Exec finished

Tools

Quick Execution

Tool Description
run Create (or reuse) a terminal and execute a command in one step. Returns clean output with exit code.

Session Management

Tool Description
create Create a new visible terminal session. Returns a sessionId.
exec Execute a command in an existing session and capture output.
read Read output from a session with pagination. Supports incremental reads and tail mode (offset: -N).
input Send text to an interactive terminal (prompts, REPLs, confirmations).
list List active sessions. Optionally filter by agentId.
close Close a terminal session and its VSCode tab.

Usage Patterns

Simple Command

The run tool handles everything — creates a terminal if needed, executes, and returns clean output:

> Run npm test
$ npm test
PASS src/utils.test.ts (3 tests)
PASS src/index.test.ts (5 tests)

[exit: 0 | 1243ms | session-abc123]

Long-Running Process

For builds, deployments, or any command that takes a while:

> Start `npm run build` without waiting, then check progress

The agent will:

  1. Call run with waitForCompletion: false — returns immediately
  2. Call read with offset: -10 to check the last 10 lines
  3. Repeat until the process completes

Interactive Commands

For commands that need user input:

> Run npm init and answer the prompts

The agent will:

  1. Call run with npm init
  2. Call read to see the prompt
  3. Call input to send the answer

Parallel Agents

Subagents can work in isolated terminals using agentId:

> Have one agent run tests while another runs the linter

Each subagent gets its own terminal tagged with its agentId, preventing output from mixing.

Configuration

The extension reads configuration from VSCode settings under terminalMcp.*:

Setting Type Default Description
terminalMcp.maxSessions number 10 Maximum concurrent terminal sessions
terminalMcp.commandTimeout number 30000 Default command timeout in ms
terminalMcp.maxOutputLines number 5000 Max lines kept in output buffer per session
terminalMcp.idleTimeout number 1800000 Close idle sessions after this many ms (0 = disabled)
terminalMcp.blockedCommands string[] ["rm -rf /"] Commands that will be rejected

Recommended: Set as Preferred Tool

LLM agents like Claude Code have a built-in Bash tool that executes commands inline within the chat. The output is embedded in the conversation and hard to read, especially for commands with verbose output. We recommend telling the agent to prefer this MCP over its built-in Bash tool.

Add the following to your project's CLAUDE.md (or equivalent instructions file):

## Terminal Execution

Prefer the BashTerm MCP tools (`run`, `exec`, `read`, etc.) over the built-in Bash tool for executing commands.
BashTerm runs commands in visible VSCode terminal tabs where the user can see output in real time.
Only fall back to the built-in Bash tool for simple, non-interactive operations like reading environment variables.

For commands that may take longer than 30 seconds or produce large amounts of output (builds, test suites,
deployments, installs), use the pull mode pattern:
1. Call `run` with `waitForCompletion: false` to launch the command without blocking.
2. Call `read` with `offset: -10` to check the last 10 lines of output.
3. Repeat step 2 until you see the command has finished (look for exit messages, prompts, or "Done").
4. Report the final result to the user.

This prevents conversation timeouts and lets the user watch progress in the terminal in real time.

Why this matters:

Built-in Bash BashTerm MCP
Output visibility Embedded in chat, hard to scroll Visible in VSCode terminal tab
Real-time feedback User sees nothing until command finishes User watches output live
Long-running commands Blocks the conversation until timeout Fire-and-forget + polling
Session state Each command is isolated Persistent sessions with history
Interactive commands Not supported Send input to prompts/REPLs

Development: Updating the Extension

VSCode aggressively caches extensions in memory. When developing locally, code --install-extension and even "Developer: Reload Window" may not reload your changes. Use this workflow:

Quick update (no restart needed)

After modifying source files, build and copy directly into the installed extension directory:

cd /path/to/vscode-terminal-mcp
npm run build
cp dist/extension.js ~/.vscode/extensions/sirlordt.vscode-terminal-mcp-<version>/dist/extension.js

Then run "Developer: Reload Window" (Ctrl+Shift+P).

Full reinstall (when quick update doesn't work)

If VSCode still uses old code:

# 1. Uninstall and remove all copies
code --uninstall-extension sirlordt.vscode-terminal-mcp
rm -rf ~/.vscode/extensions/sirlordt.vscode-terminal-mcp-*

# 2. Check for ghost entries with old publisher names
# Look in ~/.vscode/extensions/extensions.json for stale entries
# Remove any entries with old publisher IDs (e.g., "terminal-mcp.vscode-terminal-mcp")

# 3. Close VSCode completely (not just reload)

# 4. Rebuild and install
npm run build
npx vsce package --allow-missing-repository
code --install-extension vscode-terminal-mcp-<version>.vsix --force

# 5. Open VSCode

Verify the correct version is loaded

# Check which extension directories exist
ls ~/.vscode/extensions/ | grep terminal

# Verify your changes are in the installed extension
grep "YOUR_UNIQUE_STRING" ~/.vscode/extensions/sirlordt.vscode-terminal-mcp-*/dist/extension.js

# Compare checksums
md5sum dist/extension.js ~/.vscode/extensions/sirlordt.vscode-terminal-mcp-*/dist/extension.js

Large Output Handling

When read returns output that exceeds the MCP client's token limit, the system automatically saves the full output to a temporary JSON file and returns the file path in the error message.

To extract the relevant content:

# Get the last 50 lines (most relevant for status)
tail -50 /path/to/saved/file.txt

# Or parse the JSON to extract the text content
python3 -c "import json; data=json.load(open('/path/to/file.txt')); print(data[0]['text'][-2000:])"

The file format is JSON: [{"type": "text", "text": "..."}]

This commonly happens with commands that produce heavy TUI output (progress bars, ANSI escape codes). Use smaller offset values (e.g., offset: -20 instead of offset: -100) to reduce the captured output size.

How It Works

  1. The VSCode extension activates and starts an IPC server on a Unix socket
  2. The MCP entry point (mcp-entry.js) is spawned by the MCP client and bridges JSON-RPC stdio with the IPC socket
  3. Commands execute in real VSCode terminals using the Shell Integration API for reliable output capture and exit code detection
  4. Output is stored in circular buffers with pagination support for efficient reading

Latest Changes (0.1.6)

  • Screenshots in README for marketplace
  • Clean output format for all tools — no more raw JSON
  • Fixed waitForCompletion: false not working
  • Disabled idle reaper — user closes sessions manually
  • Unique IPC socket per workspace (multi-instance support)
  • Custom terminal tab names with date format
  • Large output handling documentation

See CHANGELOG.md for full history.

License

MIT

推荐服务器

Baidu Map

Baidu Map

百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。

官方
精选
JavaScript
Playwright MCP Server

Playwright MCP Server

一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。

官方
精选
TypeScript
Magic Component Platform (MCP)

Magic Component Platform (MCP)

一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。

官方
精选
本地
TypeScript
Audiense Insights MCP Server

Audiense Insights MCP Server

通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。

官方
精选
本地
TypeScript
VeyraX

VeyraX

一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。

官方
精选
本地
graphlit-mcp-server

graphlit-mcp-server

模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。

官方
精选
TypeScript
Kagi MCP Server

Kagi MCP Server

一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。

官方
精选
Python
e2b-mcp-server

e2b-mcp-server

使用 MCP 通过 e2b 运行代码。

官方
精选
Neon MCP Server

Neon MCP Server

用于与 Neon 管理 API 和数据库交互的 MCP 服务器

官方
精选
Exa MCP Server

Exa MCP Server

模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。

官方
精选