Install in four commands, then decide what the agent may touch
These docs cover installation, client configuration, the policy model and the audit format. Per-integration tool lists live on each integration page — this page is the shared layer underneath all of them.
Quickstart
Pre-release. The commands on this page describe the interface we are shipping. Nothing is published to npm or PyPI yet — get notified at launch.
Pick an integration, install the native plugin, start the MCP server, and register it with your client. This example uses Blender; the shape is identical everywhere.
blender --command extension install dccmcp_blender
npx dccmcp-blender@latest serve --port 7331
{
"mcpServers": {
"blender": {
"command": "npx",
"args": ["dccmcp-blender", "serve"]
}
}
}dccmcp doctor --tool scene.inspect
The default policy is read-only. Write tools stay blocked until you approve them, which is why a first session can safely be exploratory.
Install matrix
| Software | Plugin / package | Server |
|---|---|---|
| Blender | blender --command extension install dccmcp_blender | npx dccmcp-blender serve |
| Maya | Maya module (mod) in MAYA_MODULE_PATH | npx dccmcp-maya serve |
| Houdini | Houdini package in HOUDINI_PACKAGE_PATH | npx dccmcp-houdini serve |
| 3ds Max | 3dsmaxbatch -U dccmcp-3dsmax.mzp | npx dccmcp-3dsmax serve |
| Rhino | yak install dccmcp-rhino | npx dccmcp-rhino serve |
| ZBrush | ZScript dropped into ZStartup | npx dccmcp-zbrush serve |
| Photoshop | uxp install ./dccmcp-photoshop.ccx | npx dccmcp-photoshop serve |
| FreeCAD | Addon Manager → DCCMCP Workbench | npx dccmcp-freecad serve |
| QGIS | Plugins → Manage and Install → DCCMCP | npx dccmcp-qgis serve |
| OpenCV | pip install dccmcp-opencv | dccmcp-opencv serve |
Agent clients
Every server speaks the Model Context Protocol over stdio and streamable HTTP, so the client you already use is the client you keep. These are the clients we test each release; if your agent is not listed but speaks MCP, it works.
- Claude CodeAnthropic
- CursorAnysphere
- CodexOpenAI
- GitHub CopilotGitHub
- Claude DesktopAnthropic
- Gemini CLIGoogle
- WindsurfCognition
- ClineOpen source
- Roo CodeOpen source
- TraeByteDance
- DoubaoByteDance
- Qwen CodeAlibaba
- KimiMoonshot AI
- Zhipu GLMZhipu AI
- CodeBuddyTencent
- YuanbaoTencent
- LingmaAlibaba
- QoderAlibaba
- ContinueOpen source
- ZedZed Industries
- JetBrains AIJetBrains
- GooseBlock
- WarpWarp
- CozeByteDance
- CodeGeeXZhipu AI
- ComateBaidu
- ERNIEBaidu
- SparkiFlytek
- MiniMax AgentMiniMax
- OpenClawOpen source
- WorkBuddyIndependent
- Your own agentMCP SDK
{
"mcpServers": {
"blender": { "command": "npx", "args": ["dccmcp-blender", "serve"] }
}
}{
"mcpServers": {
"blender": {
"type": "http",
"url": "http://studio-host:7331/mcp",
"headers": { "Authorization": "Bearer <token>" }
}
}
}Clients that require a command and arguments (rather than a URL) use the stdio entry. Clients running on another machine use the HTTP entry, which binds to loopback unless you explicitly widen it.
Tool reference
Each server publishes its tool list at connect time. The full list, with arguments and risk classes, is documented per integration.
Policies & approvals
A policy decides what happens when a tool is called: run it, ask a human, or refuse. Policies are declarative files, so they can be reviewed, versioned and shared across a team.
[default]
mode = "read-only"
[[rule]]
tool = "object.create"
action = "require-approval"
[[rule]]
tool = "pipeline.checkpoint.restore"
action = "deny"
when = { path = "**/production/**" }
[[rule]]
tool = "scene.inspect"
action = "allow"
audit = "summary"Approvals can be interactive — the agent pauses and a human confirms in the client — or asynchronous, where the request is queued for a reviewer and the agent continues with other work.
Audit & rollback
Every call writes a structured entry. Entries are append-only on disk and can be streamed to a webhook or SIEM.
2026-09-20T09:14:22Z studio/leon
tool geometry_nodes.set_input
args { "socket": "Inset", "value": 0.02 }
result ok · 118 ms · 3 variants
file facade_module_01.blend @ sha256:9f2c…
host Blender 4.5.1dccmcp restore --checkpoint facade-a1.ckpt --scope objects
Skills SDK
Register your own tools when studio conventions matter. A skill is a typed callable plus a manifest, and it inherits the same policy and audit behavior as built-in tools.
from dccmcp import tool, Skill
skill = Skill(
name="studio-naming",
tools=[
tool(
name="naming.rename_collection",
risk="write",
schema={"collection": "str", "pattern": "str"},
)(rename_collection),
],
)Changelog
- v1.4.02026-09-08
- Added streamable HTTP transport with token auth for LAN deployments.
- Blender: geometry node simulation zones are now addressable by socket path.
- QGIS: processing provenance now records algorithm provider and version.
- v1.3.22026-08-19
- Rhino: block instance transforms accept rotation constraints.
- Fixed a checkpoint naming collision when two sessions wrote the same file.
- v1.3.02026-08-02
- Skills SDK 1.0 — register custom typed tools from Python or .NET.
- Audit entries now include a host application version fingerprint.
Status & support matrix
| Software | Supported versions | Update cadence |
|---|---|---|
| Blender | 4.2 LTS · 4.5 LTS · 5.x | Tested within 14 days of a vendor release |
| Maya | 2024 · 2025 · 2026 | Tested within 14 days of a vendor release |
| Houdini | 20.5 · 21+ | Tested within 21 days of a stable release |
| 3ds Max | 2024 · 2025 · 2026 | Tested within 14 days of a vendor release |
| Rhino | Rhino 7 · 8 · 9 WIP track | Tested within 14 days of a vendor release |
| ZBrush | 2024 · 2025+ (desktop) | Tested within 30 days of a stable release |
| Photoshop | 2024 · 2025 · 2026 | Tested within 14 days of a vendor release |
| FreeCAD | 0.21 · 1.0+ | Tested within 21 days of a stable release |
| QGIS | 3.34 LTR · 3.40+ | Tested within 14 days of a LTR release |
| OpenCV | 4.x · 5.x | Tracked continuously |
Deploying for a team?
We will walk your IT or security team through transport, storage and policy export.
Looking for the security posture instead? Read the security overview.