RapidMiner / Altair AI Studio MCP
# RapidMiner / Altair AI Studio MCP
An MCP server that lets Claude open, see, and control **Altair AI Studio** (the current name for RapidMiner Studio).
```
Claude ──MCP/stdio──▶ src/index.js (Node) ──HTTP 127.0.0.1:8765 + token──▶ Java agent inside AI Studio's JVM
├─ Swing UI control (screenshots, clicks, typing, menus, trees…)
└─ RapidMiner API (process, operators, parameters, run, results, settings)
```
The agent runs **inside** AI Studio (loaded with `-javaagent`). That way:
- screenshots and clicks work on **GNOME Wayland**, with no xdotool, ydotool or screen-capture permissions;
- Claude reads the real component tree (buttons, trees, tables, dialogs) instead of guessing pixels;
- processes can be built and configured exactly through RapidMiner's own API, with Groovy as an escape hatch.
## Setup
```bash
npm install
./agent/build.sh # compiles agent/aistudio-mcp-agent.jar against your AI Studio jars
claude mcp add rapidminer -s user -- node "$PWD/src/index.js"
```
### Using with OpenCode
Add the server to `~/.config/opencode/opencode.jsonc` (or an `opencode.json` in your project), using the absolute path to `src/index.js`:
```jsonc
{
"$schema": "https://opencode.ai/config.json",
"mcp": {
"rapidminer": {
"type": "local",
"command": ["node", "/absolute/path/to/AIStudioMCP/src/index.js"],
"enabled": true
// "environment": { "RAPIDMINER_HOME": "/path/to/AI Studio" }
}
}
}
```
AI Studio is expected at `~/Downloads/Altair/RapidMiner/AI Studio 2026.1.1`. If yours is elsewhere, set `RAPIDMINER_HOME`
(for both `agent/build.sh` and the MCP server's environment).
Start AI Studio **through the launcher** (`npm run launch`, or let Claude call `rm_launch`). The launcher matches the official
`AI-Studio.sh` and only adds the agent. Your installation is not modified. If you start AI Studio from its normal shortcut, the
agent is not loaded and the tools can't connect.
## Tools
| Group | Tools |
|---|---|
| Lifecycle | `rm_launch`, `rm_status` |
| Screen control | `rm_windows`, `rm_screenshot`, `rm_ui_tree`, `rm_ui_find`, `rm_ui_get`, `rm_click`, `rm_type`, `rm_key`, `rm_select`, `rm_menu`, `rm_scroll`, `rm_drag`, `rm_window` |
| Process | `rm_process_structure`, `rm_process_xml`, `rm_process_set_xml`, `rm_process_new`, `rm_process_open`, `rm_process_save`, `rm_process_validate`, `rm_process_run`, `rm_process_stop`, `rm_process_results` |
| Operators | `rm_operator_search`, `rm_operator_info`, `rm_operator_add`, `rm_operator_get`, `rm_operator_set_parameters`, `rm_operator_remove`, `rm_operator_rename`, `rm_operator_enable`, `rm_operator_select`, `rm_connect`, `rm_disconnect`, `rm_arrange` |
| Configuration | `rm_settings_get`, `rm_settings_set`, `rm_log`, `rm_groovy` |
Tips
- Find components by `id` (from `rm_ui_tree`/`rm_ui_find`) or by `text` / `name` / `tooltip` / `class` (class matches superclasses too, e.g. `JTree`).
- `rm_screenshot` returns window coordinates scaled by `scale`. Pass the same `scale` to `rm_click`/`rm_drag` when clicking by `x`,`y`.
- Some screens (e.g. the Welcome dialog) are custom-painted: use a screenshot plus an `x`,`y` click.
- Port notation for `rm_connect`: inner ports of a subprocess are `<Chain>:inner` (the main process is `Process:inner` or `process`).
- `rm_window close` needs an explicit window id. Closing the main window also needs `confirm_exit: true`.
## Files
- `src/index.js`: MCP server (Node, stdio)
- `agent/src/aistudiomcp/*.java`: the Java agent (`Agent` HTTP server, `Ui` Swing control, `Proc` RapidMiner API, `LogCapture`)
- `scripts/launch-ai-studio.sh`: starts AI Studio with the agent (it also fixes missing `DISPLAY`/`XAUTHORITY` and the UTF-8 locale when started from MCP)
- `examples/iris_decision_tree.rmp`: a sample process built entirely through the MCP
## Security
The agent listens on `127.0.0.1` only. Every request needs a random token that is regenerated each launch and written to
`~/.AltairRapidMiner/aistudio-mcp.json` (mode 600). `rm_groovy` can run arbitrary code inside AI Studio, so treat this MCP
like shell access.
TDQS
Scored across 41 tools
Most tools are distinguishable by domain and action, but several clusters have close boundaries: rm_ui_tree/rm_ui_find/rm_ui_get, rm_click/rm_select/rm_menu, and rm_window/rm_windows. Descriptions help, but an agent must read them carefully to choose correctly, so overlap remains notable.
All names use consistent snake_case with an rm_ prefix, and most follow domain_action patterns such as rm_process_run and rm_operator_add. There are minor deviations with verb-only or noun-only names like rm_click, rm_type, rm_window, and rm_windows, but the overall convention is readable.
At 41 tools, the surface is far larger than the typical well-scoped range and includes many granular UI-automation and process-introspection operations. Several tools could be consolidated, making the set feel heavy for an agent to navigate.
The server covers process lifecycle, operator management, UI interaction, settings, logs, and even a Groovy escape hatch, so major workflows are supported. Minor gaps remain around repository/data browsing and explicit process close/delete operations, but these are mostly workable via existing tools.