pyside6-mcp
<!-- mcp-name: io.github.com55/pyside6-mcp -->
# pyside6-mcp
Playwright-style MCP server for PySide6 apps — lets AI assistants see, control, and debug your Python desktop GUI without modifying your app's source code.
```
AI assistant → MCP tools → pyside6-mcp server → HTTP bridge (ephemeral localhost port) → PySide6 app
```
## Features
- **Screenshot** any window or widget
- **Inspect** the full widget tree (class, name, geometry, text, state)
- **Click, type, scroll, press keys** — full interaction
- **Find widgets** by class, objectName, or text content
- **Read Python logs** captured from the app
- **Run Python** inside the app process for advanced inspection
- **Launch and stop** the app from the agent
Zero changes to your app's source code required.
## Requirements
- Python 3.11+
- PySide6 6.6+
- [uv](https://docs.astral.sh/uv/) (recommended)
- Windows (tested), Linux/macOS (should work)
## Installation
The MCP server (stdio) and the in-app bridge are separate:
| Component | Where it runs | Needs PySide6? |
|-----------|---------------|----------------|
| **MCP server** (`pyside6-mcp`) | MCP client's process | No |
| **Bridge** (`python -m pyside6_mcp …`) | Inside your PySide6 app | Yes |
**No changes to the target project.** `launch_app(cwd=…)` injects only the
bridge modules via an isolated `PYTHONPATH` — nothing is added to `pyproject.toml`,
and the MCP server's dependencies (fastmcp, httpx, …) never enter the app environment.
Requires [uv](https://docs.astral.sh/uv/) on `PATH`. MCP Registry name:
`io.github.com55/pyside6-mcp`.
Unlike [qt-mcp](https://github.com/0xCarbon/qt-mcp), you do **not** add this
package to the target app or set a probe env var. `launch_app` injects the bridge.
### Claude Code
Plugin (MCP + skill). `install` only searches marketplaces you have added —
there is no `github:` install shorthand:
```bash
claude plugin marketplace add com55/pyside6-mcp
claude plugin install pyside6-mcp@pyside6-mcp
```
MCP server only:
```bash
claude mcp add -s user pyside6 -- uvx pyside6-mcp
```
### Cursor
Install the Agent Plugin from this repository (or from the Cursor Marketplace
once listed). Manual stdio config:
```json
{
"mcpServers": {
"pyside6": {
"command": "uvx",
"args": ["pyside6-mcp"]
}
}
}
```
After saving, reload MCP servers in **Cursor Settings → MCP**.
See [`examples/cursor-mcp-config.json`](examples/cursor-mcp-config.json).
### Any stdio MCP client
```json
{
"mcpServers": {
"pyside6": {
"command": "uvx",
"args": ["pyside6-mcp"]
}
}
}
```
Place it wherever that client expects MCP config (user-level or project-level).
See [`examples/mcp-config.json`](examples/mcp-config.json).
### VS Code / GitHub Copilot / other clients
Same stdio block as **Any stdio MCP client**.
> **Note:** `uvx` downloads and runs the MCP server in an isolated env — PySide6 is **not**
> required there. PySide6 is only needed in the target app's venv (already a project dependency).
>
> To run from a git checkout instead of PyPI: `uvx --from git+https://github.com/com55/pyside6-mcp pyside6-mcp`
## Usage
### From MCP (recommended — zero project setup)
```
launch_app(cwd="/path/to/project") # main.py at root
launch_app(cwd="/path/to/project", script="app.py") # other name at root
launch_app(cwd="/path/to/project", script="backend/gui.py") # entry in subfolder
get_launch_help() # full script decision guide
list_apps() # pids of apps this server launched
```
`cwd` is the project root (`pyproject.toml`). `script` is the entry `.py` **relative to cwd**.
`launch_app` returns `{pid}` (the Qt app process, not the `uv` wrapper). Other tools omit `pid` to target the last launched app; pass `pid=` when several are running. Do not pass `port=`.
### Manual launch (when you start the app yourself)
Prefer `launch_app` above. Use these only when the app is already running, or
you want to start it outside the MCP server.
**Without editing app source** — run the entry script through the bridge launcher
(pulls `pyside6-mcp` into a temporary env; does not add it to the project):
```bash
cd your-pyside6-project
uv run --with pyside6-mcp python -m pyside6_mcp main.py
# other entry: … python -m pyside6_mcp app.py
# in a subfolder: … python -m pyside6_mcp backend/gui.py
```
Bridge listens on `http://127.0.0.1:7890` by default. Override with
`PYSIDE6_MCP_PORT`. Then from the agent call `wait_until_ready()` (and pass
`pid=` only if you use several apps).
`launch_app` is different: it picks an ephemeral port and injects only the
bridge modules via an isolated `PYTHONPATH` — you never set the port yourself.
**Embed in the app** (always-on; requires adding the package to that project):
```bash
uv add pyside6-mcp
```
```python
# In your app's main(), after QApplication is created and before app.exec()
from pyside6_mcp import install_bridge
install_bridge() # same default port 7890 / PYSIDE6_MCP_PORT
```
### From your AI assistant
Once the app is running with the bridge active, ask your assistant:
> "Screenshot the app and click the Apply button"
> "Why is the checkbox disabled? Inspect its state."
> "Fill in the form and submit it"
> "Show me the last 20 log lines from the app"
Your assistant uses the `launch_app`, `screenshot`, `get_widget_tree`, `find_widget`, `click`, `type_text`, `get_logs`, and other tools automatically.
`launch_app` returns only when the UI is ready (a visible top-level window that has been quiet for at least 500 ms), not merely when the bridge HTTP server is up. Default `timeout` is 45 seconds.
## Tools
| Tool | Description |
|------|-------------|
| `launch_app(cwd, script?, app_args?, timeout)` | Launch app (no target-project install) and wait for UI readiness. Returns `{pid}` |
| `list_apps()` | Pids of apps launched by this server |
| `get_launch_help()` | How to set cwd, script, app_args before launch |
| `wait_until_ready(timeout, quiet_ms, pid?)` | Wait for UI readiness on an already-running app |
| `wait_for_idle(timeout, quiet_ms, pid?)` | Wait until UI has been quiet after an action |
| `get_app_status(pid?)` | Process + bridge health; detects likely modal blocks |
| `stop_app(pid?)` | Stop a launched app |
| `screenshot(widget_id?, pid?)` | Capture window or specific widget (modal/active-window aware) |
| `get_widget_tree(pid?)` | Full widget hierarchy with IDs |
| `get_widget_info(widget_id, pid?)` | Detailed properties of one widget |
| `get_app_state(pid?)` | Active window, focus, screen info |
| `find_widget(class_name?, object_name?, text?, visible?, pid?)` | Search widgets |
| `click(widget_id?, x?, y?, button?, pid?)` | Mouse click |
| `double_click(widget_id, x?, y?, pid?)` | Double click |
| `type_text(text, widget_id?, pid?)` | Keyboard input |
| `press_key(key, pid?)` | Named key: enter, escape, tab, up/down, f5, … |
| `scroll(dy, widget_id?, dx?, pid?)` | Scroll wheel |
| `list_actions(pid?)` | List QAction menu/toolbar items |
| `trigger_action(name?, text?, pid?)` | Trigger a QAction without clicking menus |
| `get_logs(n?, pid?)` | Recent Python log records |
| `get_app_output(n?, pid?)` | Raw stdout/stderr from launched app |
| `eval_python(code, pid?)` | Execute Python inside the app process |
## Agent Skill
A companion skill ships at [`skills/pyside6-mcp/SKILL.md`](skills/pyside6-mcp/SKILL.md)
and is installed automatically with the plugin. It tells the agent when and how
to use these tools — no need to explain the workflow every time.
## Architecture
```
pyside6_mcp/
├── bridge.py # In-process HTTP server (runs inside the target app)
├── launch.py # Isolated PYTHONPATH inject + launch_app argv/validation
├── session.py # Handshake + in-memory pid → port sessions
├── server.py # FastMCP stdio server (the MCP client talks to this)
├── __init__.py # Exports install_bridge()
└── __main__.py # Launcher: python -m pyside6_mcp <script>
```
**Thread safety**: all Qt operations are marshaled to the main thread via `QApplication.postEvent` with a custom event type — the same mechanism Qt uses internally for cross-thread signals.
## Examples
- [`examples/test_app.py`](examples/test_app.py) — minimal PySide6 app to verify the bridge
- [`examples/mcp-config.json`](examples/mcp-config.json) — generic MCP config (Cursor, VS Code, Windsurf, …)
- [`examples/cursor-mcp-config.json`](examples/cursor-mcp-config.json) — Cursor-specific template
- [`examples/claude-mcp-config.json`](examples/claude-mcp-config.json) — same format, kept for reference
## License
MIT
TDQS
Scored across 22 tools
Most tools target distinct resources/actions: screenshot/tree/info/find for inspection, click/type/press/scroll for input, launch/stop/status for lifecycle. A couple of pairs could be confused: wait_until_ready vs wait_for_idle and get_app_state vs get_app_status, though their descriptions separate them.
The set predominantly uses snake_case verb phrases such as get_widget_tree, list_actions, launch_app, and trigger_action. Minor deviations like the noun command 'screenshot' and a few bare verbs (click, scroll) keep it from being perfectly uniform, but the overall pattern is predictable.
22 tools is on the heavy side for an MCP surface, and some functionality (wait_until_ready/wait_for_idle, get_app_state/get_app_status, get_logs/get_app_output) could potentially be consolidated. Still, the count is defensible given the breadth of UI automation and debugging features.
The set covers the full lifecycle of launching, inspecting, interacting with, waiting on, and stopping a PySide6 app, plus logs and QAction access. There are minor gaps such as no direct property-mutation or drag/drop tool, but eval_python and existing input tools provide workarounds.