Skip to content
BetaFX & Simulation

MCP for Houdini — Connect AI Agents to Houdini

Procedural work is a graph of intent. Agents get typed access to that graph — nodes, parameters, expressions, VEX — with cooks scoped so a single question never triggers an eight-hour simulation.

Read-only by default12 typed toolsCheckpointed writes
Technical specification
Host software
Houdini 20.5, 21 and newer (Core, FX, Indie)
Runtime
Python 3.11 via hython and the Houdini Object Model
Transport
stdio (local) and streamable HTTP (farm/submitter)
Install
Houdini package + npx dccmcp-houdini
Systems
SOPs, VEX, HDAs, PDG/TOPs, USD, Karma, Mantra
Platforms
Windows, Linux, macOS
Last reviewed
2026-09-18
Overview

What is MCP for Houdini?

MCP for Houdini is a local MCP server plus a Houdini plugin that connects AI agents to a real Houdini session. Agents discover 12 typed tools — including network.inspect, parm.inspect, node.create and node.connect — and call them to inspect, create, modify and validate work inside the live application.

Unlike generated Houdini scripts, every call is validated against your current document, classified as read, write or execute, and recorded in an audit log. Houdini stays open, the document stays live, and destructive operations are checkpointed so they can be rolled back.

It works with Claude Code, Codex, Cursor, GitHub Copilot, Claude Desktop, Gemini CLI, Trae, Doubao, Qwen Code, CodeBuddy, Custom MCP clients, Farm submitters over stdio or streamable HTTP.

  • Host software: Houdini 20.5, 21 and newer (Core, FX, Indie)
  • Runtime: Python 3.11 via hython and the Houdini Object Model
  • Transport: stdio (local) and streamable HTTP (farm/submitter)
Why it breaks

What goes wrong when agents drive Houdini without a contract

01

One parameter change can trigger an eight-hour cook

Procedural networks are only cheap until something invalidates upstream. An agent poking at random parameters can queue a simulation nobody asked for.

02

VEX written blind is VEX that does not compile

Generated snippets fail on type mismatches, missing attributes or the wrong run-over context — and the error surfaces deep in a cook rather than at the point of authorship.

03

TOPs graphs are powerful and opaque

Work items, dependencies and attributes are exactly the kind of structured state agents are good at, but only if the graph can be inspected rather than inferred.

How it works

How MCP for Houdini keeps intent, structure and traceability intact

An MCP server and Houdini package that exposes node networks, parameters, VEX snippets, digital assets, PDG/TOPs graphs and USD/Karma output as typed, checkpointed tools.

Scoped cooking

Every operation declares what may cook. The agent gets a progress stream and a cancel handle instead of a blocked session, and nothing upstream is invalidated implicitly.

Parameter binding, not hardcoding

Tools accept literal values or channel references, and the server reports which parameters remain hardcoded so procedurally intended setups stay procedural.

VEX with validation

Snippets are linted for run-over context, attribute types and missing bindings before they are written to the node, and the compile error comes back as a structured result.

PDG as a typed surface

The agent inspects work items, sets attributes and dependencies, runs a scoped subset of the graph, and reads back the failures — the same loop a TD runs by hand.

Tool reference

MCP for Houdini tools — every tool and its risk class

Agents discover this list at connect time. You decide which entries are auto-approved, which need a human and which are denied in production files.
  • network.inspectNode network structure, node types, flags, errors and warnings.read
  • parm.inspectParameter values, channel references, expressions and defaults.read
  • node.createCreate nodes in an explicit parent with validated type names.write
  • node.connectWire inputs and outputs with cycle detection before the edit lands.write
  • parm.setSet parameter values by component, with optional channel references.write
  • parm.expressionSet and evaluate HScript or Python expressions in context.write
  • vex.set_snippetAuthor a VEX snippet with run-over context validation and a compile check.write
  • hda.instantiateInstantiate digital assets and set typed, versioned parameter sets.write
  • geometry.cookCook a scoped node, stream progress and return geometry statistics.execute
  • pdg.runRun a scoped subset of a TOP graph and return per-work-item status.execute
  • usd.exportExport USD layers with explicit stage, root prim and unit settings.execute
  • cache.writeWrite geometry, volumes or simulation caches to declared paths.execute
Install

Install MCP for Houdini in four steps

Everything runs locally. The server binds to loopback by default and only exposes the tools you approve in Houdini.

Pre-release. These commands describe the interface we are shipping for MCP for Houdini — the packages are not published yet. Get notified when MCP for Houdini launches.

STEP 01

Install the Houdini package

Add the DCCMCP package to your Houdini search path. Package files keep studio and project installs separate.

export HOUDINI_PACKAGE_PATH=/studio/tools/dccmcp:$HOUDINI_PACKAGE_PATH
STEP 02

Start the bridge

Attach to a running session, or use hython for batch cooks on a farm node.

npx dccmcp-houdini@latest serve --mode attach
STEP 03

Register the client

Register the server with Claude Code, Codex, Cursor, OpenClaw or any other MCP client. Farm workers use the same entry with the HTTP transport.

{
  "mcpServers": {
    "houdini": { "command": "npx", "args": ["dccmcp-houdini", "serve"] }
  }
}
STEP 04

Cap the cooks

Set cook time and memory ceilings so an agent cannot accidentally launch a full simulation on a workstation.

dccmcp policy limits --cook-seconds 300 --memory 12Gi

MCP for Houdini works with the agents you already use

  • Claude Code
  • Codex
  • Cursor
  • GitHub Copilot
  • Claude Desktop
  • Gemini CLI
  • Trae
  • Doubao
  • Qwen Code
  • CodeBuddy
  • Custom MCP clients
  • Farm submitters

Any other MCP-compatible client works too — one config entry, stdio by default.

Read the docs
FAQ

Questions teams ask before they install

If something is missing, ask us directly — we answer technical questions with technical answers.
  • No. Cooking is an explicit tool call with a scope, and a policy can cap cook time and memory. Reading node structure and parameter values never triggers a cook.

  • Yes. The PDG graph is exposed as structured data — work items, attributes, dependencies and results — so an agent can inspect, run a scoped subset and report failures per work item.

  • It can author VEX snippets and parameter expressions, both validated before they are committed to the node. Invalid snippets come back as structured compile errors the agent can correct.

  • Houdini 20.5 and 21 and newer, across Core, FX and Indie licenses. Indie users get the same tool surface; parallel cook limits follow the license you have installed.

  • The same tools run under hython on a farm node, and caches, USD exports and renders are declared as paths your submitter already understands.

npx dccmcp-blender@latest serve

Put an agent inside your real tools — without putting your pipeline at risk

Install an integration, set a policy, and let your agent inspect, build and validate work in Blender, Maya, Houdini, 3ds Max, Rhino, ZBrush, Photoshop, FreeCAD, QGIS or OpenCV. Free to start, local by default.

macOS · Windows · Linux — no cloud dependency required