使用说明

MCP TUI Test

Like Playwright, but for Terminal User Interfaces

An MCP (Model Context Protocol) server that enables AI assistants to test Terminal User Interface (TUI) applications. This server provides tools to launch, interact with, and verify TUI applications programmatically.

📚 View Documentation

Features

Dual Testing Modes: Choose between stream mode (for CLI tools) or buffer mode (for full TUIs with position awareness)
Launch TUI Applications: Start any terminal-based application with configurable dimensions
Send Keyboard Input: Simulate user typing, special keys, and control combinations
Capture Screen Output: Read and analyze the current terminal display
Position-Based Testing: Verify text at specific screen coordinates (buffer mode)
Cursor Tracking: Monitor cursor position in real-time (buffer mode)
Wait for Text: Asynchronously wait for specific content to appear
Assertions: Verify that expected content is present in the output
Session Management: Run multiple TUI applications simultaneously

Testing Modes

Stream Mode (Default)

Best for: CLI tools, command-line applications, simple interactive programs
Uses: pexpect for stream-based testing
Features: Text matching, pattern waiting, output capture
Example use cases: git, npm, grep, interactive shell scripts

Buffer Mode

Best for: Full TUI applications, ncurses apps, dialog boxes, menus
Uses: pexpect + pyte for screen buffer emulation
Features: All stream features PLUS position-based assertions, cursor tracking, region extraction
Example use cases: htop, vim, dialog, interactive menus

When to use which mode:

Use stream mode for applications that output text sequentially
Use buffer mode for applications that draw complex UIs with cursor movement

Installation

Prerequisites

Python 3.10 or higher
uv

Install

Once published to PyPI, no install is needed — run it straight from uvx:

uvx mcp-tui-test

For development:

uv venv
source .venv/bin/activate
uv pip install -e ".[dev]"

Go implementation

A sibling Go implementation lives in go/ — the same tool contract (same tool names, same structured result shapes, same tui://{session_id}/screen resource), as a single static binary with zero runtime dependencies:

go install github.com/GeorgePearse/mcp-tui-test/go/cmd/mcp-tui-test@latest

{
  "mcpServers": {
    "tui-test": { "command": "mcp-tui-test" }
  }
}

Built on the official Go MCP SDK, creack/pty, and vt10x for buffer-mode screen emulation. The Python package remains the home of the pytest plugin; the two share the tool schema, so clients can't tell them apart.

Structured results

Every tool returns structured content (with an output schema), not prose: a success flag plus typed fields (screen, outcome, passed, row/col, screen_excerpt on failed assertions, error when something broke). Agents and scripts branch on fields instead of parsing ✓/✗ strings.

The current screen is also exposed as an MCP resource at tui://{session_id}/screen.

For harness-side loops, result handling, and session ordering, see Composing TUI tests in a coding harness.

Session limits & cleanup

Sessions are capped and reaped automatically:

MCP_TUI_MAX_SESSIONS (default 8) — launching past the cap returns a structured error.
MCP_TUI_SESSION_TTL_S (default 900) — sessions older than this are closed on the next access.
Dead processes are reaped on every registry access and reported by list_sessions.
All sessions are force-closed (SIGKILL escalation) at interpreter exit — no zombies.

pytest plugin

The same engine drives deterministic CI tests with no LLM in the loop. Installing the package registers a tui fixture:

def test_app_header(tui):
    s = tui.launch("my-app --demo", mode="buffer", dimensions=(120, 40))
    s.wait_until_contains("Dashboard", timeout=10)
    s.assert_at("Dashboard", 0, 2)
    s.send("q")

TuiHandle offers send, send_ctrl, expect, assert_contains, assert_at, wait_until_contains, screen, line, and cursor; failures raise AssertionError with the current screen attached. Sessions are closed at test teardown, pass or fail.

Usage

Running the MCP Server

python server.py

Or if installed as a package:

mcp-tui-test

Configure in Claude Desktop

Add this to your claude_desktop_config.json:

{
  "mcpServers": {
    "tui-test": {
      "command": "uvx",
      "args": ["mcp-tui-test"]
    }
  }
}

(A local checkout also works: "command": "python", "args": ["/path/to/mcp-tui-test/server.py"].)


## Available Tools

### `launch_tui`
Launch a TUI application for testing.

**Parameters:**
- `command` (required): The command to launch the TUI application
- `session_id` (optional): Unique identifier for this session (default: "default")
- `timeout` (optional): Command timeout in seconds (default: 30)
- `dimensions` (optional): Terminal dimensions as WIDTHxHEIGHT (default: "80x24")
- `mode` (optional): Testing mode - "stream" or "buffer" (default: "stream")

**Examples:**
```python
# Stream mode for CLI tools
launch_tui(command="python example_tui_app.py", session_id="test1")

# Buffer mode for full TUI applications
launch_tui(command="htop", session_id="test2", mode="buffer", dimensions="120x40")

`send_keys`

Send keyboard input to a TUI application.

Parameters:

keys (required): Keys to send. Use \n for Enter, \t for Tab, \x1b for Escape
session_id (optional): Session identifier (default: "default")
delay (optional): Delay in seconds after sending keys (default: 0.1)

Example:

send_keys(keys="1\n", session_id="test1")

`send_ctrl`

Send a Ctrl+Key combination to the TUI application.

Parameters:

key (required): The key to combine with Ctrl (e.g., 'c', 'd', 'z')
session_id (optional): Session identifier (default: "default")

Example:

send_ctrl(key="c", session_id="test1")

`capture_screen`

Capture the current screen output of a TUI application.

Parameters:

session_id (optional): Session identifier (default: "default")
include_ansi (optional): Whether to include ANSI escape codes in stream mode (default: False)
use_buffer (optional): Force buffer/stream mode. Auto-detects if None (default: None)

Examples:

# Auto-detect mode based on session
capture_screen(session_id="test1")

# Force buffer mode capture
capture_screen(session_id="test1", use_buffer=True)

`expect_text`

Wait for specific text to appear in the TUI output.

Parameters:

pattern (required): Text or regex pattern to wait for
session_id (optional): Session identifier (default: "default")
timeout (optional): Maximum time to wait in seconds (default: 10)

Example:

expect_text(pattern="Welcome", session_id="test1", timeout=5)

`assert_contains`

Assert that the current screen contains specific text.

Parameters:

text (required): Text to search for in the current screen
session_id (optional): Session identifier (default: "default")
use_buffer (optional): Check buffer/stream mode. Auto-detects if None (default: None)

Example:

assert_contains(text="Counter value: 1", session_id="test1")

`assert_at_position` (Buffer Mode Only)

Assert that specific text appears at a screen position.

Parameters:

text (required): Text to verify at the position
row (required): Row number (0-indexed)
col (required): Column number (0-indexed)
session_id (optional): Session identifier (default: "default")

Example:

# Verify "Error" appears at row 5, column 10
assert_at_position(text="Error", row=5, col=10, session_id="test1")

`get_cursor_position` (Buffer Mode Only)

Get the current cursor position.

Parameters:

session_id (optional): Session identifier (default: "default")

Example:

get_cursor_position(session_id="test1")
# Returns: "Cursor position (session: test1): row 10, column 25"

`get_screen_region` (Buffer Mode Only)

Extract a rectangular region of the screen.

Parameters:

row_start (required): Starting row (0-indexed, inclusive)
row_end (required): Ending row (0-indexed, exclusive)
col_start (optional): Starting column (0-indexed, inclusive, default: 0)
col_end (optional): Ending column (0-indexed, exclusive, default: end of line)
session_id (optional): Session identifier (default: "default")

Example:

# Extract rows 5-10, full width
get_screen_region(row_start=5, row_end=10, session_id="test1")

# Extract rows 5-10, columns 20-60
get_screen_region(row_start=5, row_end=10, col_start=20, col_end=60, session_id="test1")

`get_line` (Buffer Mode Only)

Get a specific line from the screen buffer.

Parameters:

row (required): Row number (0-indexed)
session_id (optional): Session identifier (default: "default")

Example:

get_line(row=3, session_id="test1")
# Returns: "Line 3 (session: test1): [line content]"

`close_session`

Close a TUI testing session.

Parameters:

session_id (optional): Session identifier (default: "default")

Example:

close_session(session_id="test1")

`list_sessions`

List all active TUI testing sessions.

Example:

list_sessions()

Example Test Scenario

Here's how you might use this MCP to test a TUI application:

Launch the application:

launch_tui(command="python example_tui_app.py")

Wait for it to load:

expect_text(pattern="Welcome to the Example TUI Application")

Interact with it:
```
send_keys(keys="1\n")
```

Verify output:

assert_contains(text="Hello, TUI Tester!")

Clean up:
```
close_session()
```

Testing the Example App

This repository includes an example TUI application (example_tui_app.py) that you can use to test the MCP server.

Run it directly:

python example_tui_app.py

Or test it through the MCP:

launch_tui(command="python example_tui_app.py")
send_keys(keys="1\n")
capture_screen()

Use Cases

Automated Testing: Verify TUI applications behave correctly
Integration Testing: Test command-line tools and interactive CLIs
Documentation: Generate screenshots and examples from TUI apps
Debugging: Inspect the state of TUI applications during development
CI/CD: Add TUI testing to your continuous integration pipeline

Technical Details

This MCP server uses:

FastMCP: For the MCP server implementation
pexpect: For spawning and controlling terminal applications
pyte: For terminal emulation and screen buffer management (buffer mode)
ScreenSession wrapper: Combines pexpect and pyte for hybrid testing

Architecture

Stream Mode: pexpect directly captures output stream
Buffer Mode: pexpect output → pyte terminal emulator → screen buffer
Auto-detection: Tools automatically use appropriate mode based on session

Limitations

Currently designed for Unix-like systems (Linux, macOS)
Windows support may require modifications (consider using winpty or similar)
Mouse support in TUIs is not currently available
Buffer mode requires slightly more memory for screen emulation
Position-based assertions only work in buffer mode

Contributing

Contributions are welcome! Please feel free to submit issues or pull requests.

License

MIT License - see LICENSE file for details

Related tools

docs/landscape.md surveys the other TUI-testing tools, what each is built on, the core ideas in each, and which of those ideas this project has or could adopt. The tools covered:

microsoft/tui-test — Rust engine with Alacritty, Ghostty, Rio and xterm.js backends; Playwright-style locators; CLI plus Rust, Python, JS and Go bindings
antithesishq/bombadil — property-based exploration of web and terminal UIs on portable-pty and libghostty-vt, with temporal properties and trace shrinking
montanaflynn/headless-terminal (ht) — Go daemon on libghostty-vt; composable --wait-* flags on send; agent skill
beorn/termless — TypeScript; twelve emulator backends with a conformance corpus; recordings replayed through the same assertions; MCP server
vyncint/termlens — Rust; predicate waits, DEC 2026 frame-complete snapshots, grid masking, answers terminal capability queries
mwunsch/termscope — Zig single binary on libghostty-vt; JSON-lines session protocol
fcoury/termwright — Rust; daemon over a Unix socket, YAML step files, box detection
charmbracelet/vhs — tape DSL to GIF/MP4 plus .ascii golden files
charmbracelet/x exp/teatest, ratatui TestBackend, Textual Pilot and pytest-textual-snapshot — in-process framework test hooks with golden or SVG snapshots
testty — Rust; semantic recipes over styles, scenarios compiled to VHS tapes
hegner123/tui-test-ghost and mjpowersjr/vterm-mcp — MCP servers on libghostty-vt, the closest in shape to this project; tui-test-ghost records tool calls to JSONL and replays them as tests
libghostty-vt and go-libghostty — the headless VT engine most of the above sit on
pexpect, pyte, libvterm — PTY stream matching and VT emulation; this project's Python implementation is pexpect plus pyte

Also: Playwright (the inspiration for this project) and the MCP specification.

Author

Created for testing TUI applications with AI assistance.

Documentation development

The site uses Material for MkDocs, matching VisionChain and Bayesian Filters.

uv venv .venv-docs
uv pip install --python .venv-docs/bin/python -r docs-requirements.txt
.venv-docs/bin/mkdocs serve
.venv-docs/bin/mkdocs build --strict

For a build you can open directly from disk with local search, use MKDOCS_OFFLINE=true .venv-docs/bin/mkdocs build --strict --site-dir /tmp/mcp-tui-test-offline. Offline builds use .html page URLs; normal builds retain the published directory URLs.

使用说明

MCP TUI Test

Features

Testing Modes

Stream Mode (Default)

Buffer Mode

Installation

Prerequisites

Install

Go implementation

Structured results

Session limits & cleanup

pytest plugin

Usage

Running the MCP Server

Configure in Claude Desktop

send_keys

send_ctrl

capture_screen

expect_text

assert_contains

assert_at_position (Buffer Mode Only)

get_cursor_position (Buffer Mode Only)

get_screen_region (Buffer Mode Only)

get_line (Buffer Mode Only)

close_session

list_sessions

Example Test Scenario

Testing the Example App

Use Cases

Technical Details

Architecture

Limitations

Contributing

License

Related tools

Author

Documentation development

相关 MCP

AKShare One MCP 股市数据服务

实时网络搜索服务

arXiv 论文检索 MCP 服务器

移动自动化MCP服务器

AI持久记忆服务器

Git仓库管理MCP服务器

`send_keys`

`send_ctrl`

`capture_screen`

`expect_text`

`assert_contains`

`assert_at_position` (Buffer Mode Only)

`get_cursor_position` (Buffer Mode Only)

`get_screen_region` (Buffer Mode Only)

`get_line` (Buffer Mode Only)

`close_session`

`list_sessions`