Cascadeur MCP Plus
An MCP server for inspecting and automating a running Cascadeur session through scene/object inspection, Python execution, native actions, and UI-level control.
Run Python inside Cascadeur on the GUI thread with
csc,app,scene,view_scene, anduiin scope, with timeout control.Query current scene details, object/behaviour/property info, selection, rig/animation data, and license flags.
Capture screenshots of windows or UI elements, restore minimized windows, and downscale images.
Inspect UI windows, element trees, dialogs, and Cascadeur action keys.
Simulate real user input: click, double/right/middle-click, drag, hover, type text, press shortcuts, and wait for UI updates.
Set QML properties or invoke QML methods/signals when clicking is not practical.
Project scene objects onto viewport coordinates for mouse interaction.
Enter Cascadeur Rig mode safely before rig generation.
Trigger Cascadeur actions by action ID like toolbar/menu commands.
Read the Cascadeur log tail to diagnose errors and Python tracebacks.
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@Cascadeur MCP Plus take a screenshot of the Cascadeur viewport and tell me what's selected"
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
English | 繁體中文
Cascadeur MCP Plus
An MCP server that lets AI agents (Claude Code, Codex, Cursor and other stdio MCP clients) animate in
Cascadeur with Cascadeur's own features: open and save scenes, set poses with numbers
and let AutoPosing move the body, key and interpolate the timeline, run AutoPhysics and cleanup, and render
videos and contact sheets with Cascadeur's own renderer. When no dedicated tool fits, it can still operate the UI
the way a person does — click, drag, type, press shortcuts, answer dialogs. Think BlenderMCP, for Cascadeur.
Cascadeur 2026.2 ships a minimal script server with a single run_script tool. MCP Plus builds on it with 27
tools: 23 group related actions (keys.set means the keys tool with "action": "set"), and run_python,
batch, screenshot and video_mocap stand alone; run_python covers run_script. It has two parts: a stdio proxy that uses only
the Python standard library, and a server that runs inside Cascadeur on its bundled Python and PySide6.
Unofficial. Not affiliated with or endorsed by Nekki. Cascadeur is a trademark of Nekki Limited.
This project contains no Cascadeur code; it talks to Cascadeur only through its Python API and Qt.
What it does
Scenes and files: new, open, save, switch and close scenes; import models and FBX animations; export FBX or GLB (
scene.open,scene.save,import_export.import_model,import_export.import_animation,import_export.export).Characters and rigs: list characters and controllers, select objects, Quick Rig a bare skeleton, retarget an animation between rigs (
characters.list,selection.select,rigging.quick_rig,rigging.retarget).Poses and AutoPosing: read and write control points in centimetres and degrees and let AutoPosing pose the rest of the body; mirror, copy and paste poses (
pose.get,pose.set,pose.mirror,autoposing.enable).Timeline and keys: animation range, keys, cycles, interpolation, IK/FK, tangents and easing (
timeline.set_range,keys.set,timeline.create_cycle,keys.interpolation,keys.ik_fk).Animation layers: create, rename, hide, lock and move objects between layers (
layers.create,layers.locked,layers.move).Physics and AI: AutoPhysics, foot-sliding and collision fixes, AI inbetweening, motion generation and experimental video mocap (
physics.enable,cleanup.fix_foot_sliding,ai_motion.inbetween,ai_motion.generate,video_mocap).Review and output: images, contact sheets, image sequences and MP4 videos from Cascadeur's own renderer and Export video window (
render.contact_sheet,render.video).Camera and display: camera presets, framing, tracking, view modes, ghosts and trajectories (
camera.preset,camera.frame,camera.track,display.view_mode).UI control, as a last resort: screenshots, the live Qt Quick element tree, dialogs, and click, drag, type and keys like a person (
screenshot,ui_inspect.tree,ui_input.click,ui_input.drag,ui_inspect.dialogs).Python, jobs and undo: run Python with
csc,appandscenein scope, run several calls in onebatch, follow background jobs, undo and redo, trigger actions and read the log (run_python,jobs.get,history.undo,actions.call,status.log).Version check:
status.statuslists the actions this Cascadeur build cannot run; calling one returns an explicit error instead of failing halfway.
Related MCP server: MaxForge MCP
Requirements
Windows 10/11
Cascadeur 2026.2 with its built-in MCP script server (tested with 2026.2.3)
Python 3.11+ for the proxy and the CLI (standard library only — nothing to
pip install)An MCP client that runs stdio servers
Details: Installation.
Quick start
Get the code:
git clone https://github.com/Karylcode/cascadeur-mcp-plus.gitRegister the proxy with your MCP client. From the checkout, one command registers it with every client it detects (Claude Code, Codex CLI, Cursor, VS Code) and backs up each config file first;
--dry-runshows the changes and writes nothing:cd C:\path\to\cascadeur-mcp-plus python -m cascadeur_mcp_plus_cli setupBy hand, Claude Code:
claude mcp add --scope user cascadeur-plus -- python -I C:\path\to\cascadeur-mcp-plus\proxy.pyOther clients (JSON config):
{ "mcpServers": { "cascadeur-plus": { "command": "python", "args": ["-I", "C:\\path\\to\\cascadeur-mcp-plus\\proxy.py"] } } }Use the full path to
python.exeifpythonis not onPATH.Start Cascadeur and click Scripts → MCP → Start script server (each launch). To skip this, run
python -m cascadeur_mcp_plus_cli autostart enableonce from the checkout (one UAC prompt); Cascadeur then starts both servers itself.Ask your agent to work in Cascadeur. The first tool call starts MCP Plus inside Cascadeur automatically (about 10 seconds). After that there is no startup delay.
Next: Usage.
Troubleshooting
Something does not connect: run
python -m cascadeur_mcp_plus_cli doctorfrom the checkout. It checks the Cascadeur install, ports 8765 and 48766, both servers and the registered clients.WinError 10013on Start script server: another program owns port 8765;netstat -ano | findstr :8765shows which.Renders refuse or come out empty: Cascadeur is minimized. Restore it, for example with
screenshotandrestore=true."Unavailable in this Cascadeur version": call
status.status; it lists the missing API or action for each tool.
More: Troubleshooting.
Documentation
Documentation index: all pages, with what each one covers
Installation: requirements, setup, client configuration, autostart, environment variables
Usage: workflow, common tasks, output files, a complete example, prompt examples
Tools: all 27 tools and their actions, with parameters and defaults
How it works: proxy, in-app server, GUI-thread dispatch, background jobs, rendering
Safety: both servers run any Python they receive; read this
Behavior and limitations: what the tools cannot do, and behaviour to expect
Troubleshooting: common problems and error messages
Development: tests, linting, live tests, hot reload, source layout
Changelog: release notes (current version 0.6.0)
License
MIT. See LICENSE.
Available Tools
18 toolscall_actionB
Trigger a Cascadeur action by key via ActionManager.call_action, exactly like its toolbar button or menu item.
Inputs: key (required), settle_ms.
dialogsA
List visible dialogs/popups and secondary windows with their texts and buttons (as refs).
Inputs: no named parameters.
enter_rig_modeA
Enter Cascadeur's Rig mode using its own rig_mode.on entry points (run_after_import for a freshly imported model). Required before Quick Rigging's Add rig elements / Generate rig: without it Cascadeur crashes. Runs on the scene idle event like the official script server.
Inputs: color.
get_logB
Read the tail of Cascadeur's log file (errors and Python tracebacks land here), optionally filtered.
Inputs: lines, contains.
get_object_infoB
Describe objects by exact or partial name: type, id and behaviours with their property names.
Inputs: name (required), max_properties.
get_scene_infoB
Summarize the current scene: name/path, frames, object counts by type, selection, rig/animation behaviours, license flags.
Inputs: max_objects.
list_actionsC
List Cascadeur action keys (actionId) found in the UI, e.g. 'Commands.Rig mode.On', with labels.
Inputs: query, include_hidden.
project_pointA
Project scene objects (by name) to main-window coordinates in the 3D viewport, for ui_click/ui_drag x/y. Uses the viewport camera and a 30 degree vertical FOV (verified for Cascadeur 2026.2.3).
Inputs: names (required), fov_deg.
run_pythonA
Run Python inside Cascadeur on the GUI thread with csc, app, scene (domain scene), view_scene and ui in scope. Returns stdout and the value of a lone expression or of a variable named result. scene.modify*/edit return False when their callback failed; check it.
Inputs: code (required), timeout.
screenshotA
Capture a Cascadeur window (or one element by ref) as an image. window_units_per_pixel converts image pixels to the window coordinates used by ui_click x/y.
Inputs: ref, format, window, restore, +1 more.
ui_clickA
Click like a real user (synthesized mouse events) on an element by ref or query, or at window coordinates x/y. Supports right/middle button, double click and modifiers.
Inputs: x, y, ref, force, +7 more.
ui_dragA
Drag with the mouse like a user: press at the start, move in steps (button held), release at the end. Start: object (scene object name, projected into the 3D viewport), ref, query or window x/y. End: to_x/to_y or offset dx/dy in window coordinates. Each step runs separately so Cascadeur updates between moves (e.g. AutoPosing while dragging a control point). Returns the actual start/end points.
Inputs: x, y, dx, dy, +14 more.
ui_hoverB
Move the mouse over an element or point (shows tooltips, hover menus).
Inputs: x, y, ref, query, +2 more.
ui_invokeA
Invoke a no-argument QML method or signal on an element, e.g. 'clicked', 'toggle', 'open', 'close'. Useful when an element cannot be clicked on screen.
Inputs: ref (required), method (required).
ui_keyB
Press keys or shortcuts, space-separated combos, e.g. 'ctrl+s', 'Escape', 'ctrl+a Delete', 'F'.
Inputs: keys (required), window, settle_ms.
ui_set_propertyA
Set a QML property on an element (e.g. checked, value, text, currentIndex). Bypasses input handlers; prefer ui_click when possible.
Inputs: ref (required), name (required), value (required).
ui_treeA
List open windows and on-screen UI elements (text, tooltip, actionId, type, rect, clickable/input flags) with refs for the other ui_* tools.
Inputs: limit, query, window, include_hidden, +1 more.
ui_typeA
Type text into the focused element, or focus an element by ref/query first. replace clears it; submit presses Enter.
Inputs: text (required), ref, query, submit, +4 more.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 18 tools
Most tools have clearly distinct purposes, and the split between action control, scene introspection, and UI automation is understandable. Minor overlap exists between ui_click vs ui_invoke/ui_set_property and between dialogs vs ui_tree, but the descriptions give enough guidance to choose correctly.
Most tools follow a readable verb_noun pattern such as call_action, get_scene_info, and ui_click. However, several names are noun-only or resource-like (screenshot, dialogs, ui_tree), and the ui_* prefix mixes action tools with introspection tools, so the convention is not fully predictable.
At 18 tools the set is slightly above the typical 3-15 range, but the broad surface (actions, scene introspection, UI automation, Python execution) justifies most tools. No tool is obviously redundant, and the UI family is grouped coherently.
The set covers action discovery and execution, rig mode setup, scene/object introspection, screenshots, dialogs, and full mouse/keyboard/QML manipulation, with run_python as a broad escape hatch. Direct high-level animation or rig-generation operations are not first-class tools, but they can be composed through actions, UI automation, or Python.
Maintenance
Related MCP Connectors
Control Unreal Engine to browse assets, import content, and manage levels and sequences. Automate…
Connect AI assistants to your approved computer for files, commands, browser and desktop tasks.
Control a real browser from AI agents: navigate, click, type, screenshot, read the page.
Connect your AI assistant to Blender for 3D modeling, Geometry Nodes, materials, lighting, animation, rendering and export. Inspect real scenes and viewport images, refine the result and restore checkpoints. OAuth connects your chat to Blender Bridge Desktop on an online Windows 10/11 x64 computer running Blender 4.2+. The chat can run on another device. Requires an active Blender Bridge account; 7-day trial available. AI model access is separate.
Related MCP Servers
- AlicenseNot gradedqualityDmaintenanceEnables AI agents to control the Blockout previs desktop app for AI filmmaking, allowing staging of 3D worlds, character animation, camera framing, timeline control, and viewport screenshotting through MCP tools.7Apache 2.0
- FlicenseBqualityCmaintenanceEnables AI agents to control a live 3ds Max session directly, including creating and editing objects, materials, animation, and scene inspection, with undo support.17-
- AlicenseAqualityCmaintenanceEnables AI clients to control Cascadeur locally for scene and rig inspection, body and finger animation, keyframes, physics actions, camera control, PNG previews, and FBX import/export via stdio MCP.3610MIT
- AlicenseAqualityBmaintenanceEnables AI agents to drive Blender 5.1+ for modeling, sculpting, mesh editing, scene management, and rendering through live Python execution.261MIT