fxhoudinimcp
Provides tools for interacting with SideFX Houdini, enabling AI agents to create and manipulate 3D scenes, simulations, renderings, and more through Houdini's Python API.
Click on "Install 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., "@fxhoudinimcpcreate a sphere and add a mountain"
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.
Related MCP server: Houdini MCP Server
Table of Contents
About
A comprehensive MCP (Model Context Protocol) server for SideFX Houdini. Connects AI assistants like Claude directly to Houdini's Python API, enabling natural language control over scene building, simulation setup, rendering, and more.
188 tools, 8 resources, and 9 prompts serving 31 written workflow guides out of the box.
Features
Category | Tools | Description |
Graph Intelligence | 6 | Atomic validated network building, network verification, node doc cards, cook profiling, frame-range cooking with per-frame evidence, cook status |
Documentation | 2 | Full-text search + page retrieval over Houdini's own shipped manual (version-exact) |
Scene Management | 7 | Open, save, import/export, scene info |
Node Operations | 17 | Create, delete, copy, connect, layout, flags |
Parameters | 12 | Get/set values in bulk, expressions, keyframes, spare parameters |
Geometry (SOPs) | 14 | Points, prims, attributes, attribute statistics, volume inspection, groups, sampling, nearest-point search |
LOPs/USD | 18 | Stage inspection, prims, layers, composition, variants, lighting |
DOPs | 8 | Simulation info, DOP objects, step/reset, memory usage |
PDG/TOPs | 10 | Cook, work items, schedulers, dependency graphs |
COPs (Copernicus) | 7 | Image nodes, layers, VDB data |
HDAs | 10 | Create, install, manage Digital Assets and their sections |
Animation | 9 | Keyframes, playbar control, frame range |
Rendering | 9 | Viewport capture, render nodes, settings, render launch |
VEX | 5 | Create/edit wrangles, validate VEX code |
Code Execution | 4 | Python, HScript, expressions, env variables |
Viewport/UI | 14 | Pane management, viewer context, verified camera and renderer state, screenshots, error detection |
Scene Context | 8 | Network overview, cook chain, selection, scene summary, error analysis |
Workflows | 8 | One-call Pyro/RBD/FLIP/Vellum setup, SOP chains, render config |
Materials | 5 | List, inspect, create materials and shader networks |
CHOPs | 4 | Channel data, CHOP nodes, export channels to parameters |
Cache | 4 | List, inspect, clear, write file caches |
Takes | 4 | List, create, switch takes with parameter overrides |
Shelf Tools | 3 | Find, read and run Houdini's own shelf tools (setups build_network cannot produce) |
Architecture
flowchart LR
subgraph Client[" ๐ค AI Client "]
direction TB
A1("Claude Desktop")
A2("Cursor / VS Code")
A3("Claude Code")
end
subgraph MCP[" โก FXHoudini MCP Server "]
direction TB
B1("๐ง 188 tools")
B2("๐ฆ 8 Resources")
B3("๐ฌ 9 Prompts")
end
subgraph Houdini[" ๐ถ SideFX Houdini "]
direction TB
C1("๐ hwebserver")
C2("๐ก Dispatcher")
C3("๐๏ธ hou.* Handlers")
C1 --> C2 --> C3
end
Client -. "MCP Protocol ยท stdio" .-> MCP
MCP -. "HTTP / JSON ยท port 8100" .-> Houdini
classDef clientBox fill:#f0f4ff,stroke:#b8c9e8,stroke-width:1px,color:#2d3748,rx:12,ry:12
classDef mcpBox fill:#eef6f0,stroke:#a8d5b8,stroke-width:1px,color:#2d3748,rx:12,ry:12
classDef houdiniBox fill:#fff5f0,stroke:#e8c4a8,stroke-width:1px,color:#2d3748,rx:12,ry:12
classDef clientNode fill:#dbe4f8,stroke:#96b0dc,stroke-width:1px,color:#2d3748,rx:8,ry:8
classDef mcpNode fill:#d4edda,stroke:#82c896,stroke-width:1px,color:#2d3748,rx:8,ry:8
classDef houdiniNode fill:#fde4d0,stroke:#e0a87c,stroke-width:1px,color:#2d3748,rx:8,ry:8
class Client clientBox
class MCP mcpBox
class Houdini houdiniBox
class A1,A2,A3 clientNode
class B1,B2,B3 mcpNode
class C1,C2,C3 houdiniNodeUses Houdini's built-in hwebserver. No custom socket servers, no rpyc. Uses hdefereval.executeInMainThreadWithResult() to safely run hou.* calls on the main thread.
Installation
FXHoudini-MCP has two halves: a Houdini plugin that runs inside Houdini, and an MCP server that your AI client starts and which relays to it over loopback. Both ship in the same Python package, so one install command sets up both and one upgrade moves them together.
Requirements
Houdini 20.5+ (integration suite green on 20.5.278, 20.5.487, 20.5.613, 20.5.654, 21.0.440 and 22.0.368)
Python 3.10+, separate from the one inside Houdini
MCP SDK (
mcppackage) 1.8+, installed for you as a dependency
Install
pip install fxhoudinimcp
python -m fxhoudinimcp installThen restart Houdini, restart your MCP client, and check the MCP menu in Houdini's menu bar.
install does both halves. It writes a Houdini package file pointing at this
exact install, and registers the server with Claude Code and Claude Desktop,
whichever it finds, using the absolute path of the Python you ran it with.
Use python -m fxhoudinimcp install rather than the bare fxhoudinimcp install
if you have more than one Python. Both work, but the module form is
self-correcting: whichever interpreter runs it is the one written into your
client config, so if the command runs at all, the path it registers is correct.
Add --dry-run first if you want to see every file it would touch and change
nothing.
It asks nothing and it finishes. If you have several Houdini versions, it writes into every packages directory it finds:
Houdini plugin
Wrote C:\Users\you\Documents\houdini21.0\packages\fxhoudinimcp.json
Wrote C:\Users\you\Documents\houdini22.0\packages\fxhoudinimcp.jsonThat is safe rather than lazy. The files are identical and point at the same
plugin, so whichever directory your Houdini reads, it finds a correct one. It
also settles the Windows case where OneDrive's Documents redirection makes a
desktop-launched Houdini and a shell-launched one disagree: both paths get a
file, so both work. The cost is an MCP menu in a Houdini version you may not
use, which uninstall clears in one go.
Because it never stops to ask, the same command works unchanged from a terminal, from Houdini's MCP menu, or from a setup script. To target one directory only:
python -m fxhoudinimcp install --houdini-dir "~/Documents/houdini22.0/packages"If your MCP client already has an fxhoudini entry pointing at a different
interpreter, it is repointed at this one and the old value is printed. That is
the common case after switching Python versions or recreating a virtualenv.
Flag | What it does |
| Report every change, make none |
| Which packages directory to write into |
| Register a client, leave Houdini untouched. Needs no packages directory, so it works when several exist |
| Which client to register. |
Upgrading later moves both halves at once, because the plugin lives inside the wheel:
pip install --upgrade fxhoudinimcpThe one thing to know: if you told Houdini to load the plugin from a git
clone instead of the installed package (see by hand),
pip install --upgrade will not move that half. Those two halves are then
independent, and the server warns at startup when it finds a plugin older than
itself.
Uninstalling
pip uninstall moves neither half. The Houdini package file and the client
registration both outlive it, and both fail quietly once the package is gone: a
package file pointing at a plugin directory that no longer exists is skipped by
Houdini without a word, and a stale client entry shows up only as
"disconnected". So take the two halves out first, then the package:
python -m fxhoudinimcp uninstall
pip uninstall fxhoudinimcpuninstall lists everything it found and asks before removing any of it. Unlike
install it does not need to know which Houdini you meant: every
fxhoudinimcp.json it finds is a leftover, and the one you forget is exactly
what silently overrides your next install. Narrow it with --houdini-dir when
you only want one Houdini cleaned.
Flag | What it removes |
| Nothing. Lists what it would remove |
| Only this packages directory, instead of every one found |
| Only the client registration, leaving the package files |
| Which client to unregister from |
| Skip the confirmation. Required when stdin is not a terminal |
Configuring the plugin
The package file install writes is also where the Houdini-side settings live.
It ships every one of them at its default, so they are all visible in one place:
FXHOUDINIMCP_PORT, FXHOUDINIMCP_BIND, FXHOUDINIMCP_AUTOSTART and
FXHOUDINIMCP_AUTO_LAYOUT (see Environment Variables
for what each does). Two things to know:
Because the package sets these explicitly, it wins over the same variable set in your shell. Change them here, not in your environment. Houdini's package format has no "only if unset" method, and it rejects JSON comments, so there is no way to ship them inert.
HOUDINI_HOST,HOUDINI_PORT,MCP_TRANSPORTandLOG_LEVELdo not belong here. They are read by the MCP server process that your client launches, not by Houdini, so setting them in this file has no effect -- configure those in your MCP client instead. If you changeFXHOUDINIMCP_PORT, setHOUDINI_PORTto match on the client side.
Note that pinning HOUDINI_PORT on the client switches off the port scan. A
second Houdini moves itself to the next free port, and the client normally finds
it by scanning 8100-8115 and taking the lowest that answers. Pin it only when you
want one specific session.
Installing by hand
install is the recommended route and the rest of this section is the manual
equivalent, for contributors working from a clone, locked-down machines, or when
something needs untangling. It is the same two halves.
1. Point Houdini at the plugin
fxhoudinimcp houdini-packageThat prints the package file with the plugin path filled in for this install, plus the Houdini packages directories found on your machine. Write it with:
fxhoudinimcp houdini-package --write "~/Documents/houdini22.0/packages"Do not type the plugin path by hand. It lives inside the Python environment you
installed into, so it changes if you recreate a virtualenv, switch to uv or
pipx, or move between Python versions, and Houdini says nothing when a package
path stops resolving. --path-only prints just the path for scripting.
Like install, this deliberately does not pick a packages directory for you, and
it warns if another fxhoudinimcp.json exists elsewhere, because Houdini
processes every packages directory and lets the last one win. That is how a stale
clone silently overrides a fresh install.
Pointing at a clone instead. Contributors, or anyone wanting the plugin tracked by git, can write the package file against a checkout:
{ "env": [ { "FXHOUDINIMCP": "C:/Users/you/code/fxhoudinimcp/houdini" } ],
"path": "$FXHOUDINIMCP" }Forward slashes work on every platform. The path must end in /houdini and must
contain scripts/, MainMenuCommon.xml and the python3.Xlibs/ folders. Do not
do this and the CLI, or the two package files will fight. Remember that
pip install --upgrade cannot move a clone.
Copyinghoudini/ into your Houdini preferences directory also works, but it
is not recommended: pip cannot update a copy, so the plugin drifts behind the
server, which is the skew the startup compatibility warning exists to catch.
Use a package file so there is one copy of the plugin.
2. Point your MCP client at the server
Both examples need the absolute path to the Python that has fxhoudinimcp
installed. Clients start their servers without your shell environment, so a bare
python resolves against a PATH they may not share, and the only symptom is the
client reporting disconnected with nothing explaining why. Find the path
with:
python -c "import sys; print(sys.executable)"Claude Code (user scope, available in every project):
claude mcp add --scope user fxhoudini -- "C:\Program Files\Python311\python.exe" -m fxhoudinimcpThere is no in-place update. To repoint an existing entry, remove it first:
claude mcp remove fxhoudini -s userClaude Desktop (claude_desktop_config.json):
{
"mcpServers": {
"fxhoudini": {
"command": "C:\\Program Files\\Python311\\python.exe",
"args": ["-m", "fxhoudinimcp"]
}
}
}After any change, fully quit Claude Desktop (system tray โ Quit) and relaunch; closing the window is not enough.
To scope the server to a single project instead, add a .mcp.json in the project
root with the same mcpServers block.
python -m fxhoudinimcp install --client-only does this step for you, with the
right path already filled in, and leaves the Houdini side alone. MCP > Connect
a Client... inside Houdini prints the same command along with the port that
session actually ended up on.
When Houdini does not load the plugin
No MCP menu means the package file was skipped, and Houdini does that without printing anything. Start it with the package log enabled and look for your file:
# Windows (PowerShell)
$env:HOUDINI_PACKAGE_VERBOSE=1; houdini
# Linux / macOS
HOUDINI_PACKAGE_VERBOSE=1 houdiniA working package prints both a Loading: and a Processing: line for
fxhoudinimcp.json. Three ways this fails quietly:
A path that does not exist. Houdini skips the package and says nothing. Nothing loads: no menu, no auto-start, no
fxhoudinimcp_servermodule.A UTF-8 BOM. Houdini's JSON parser rejects a leading BOM and skips the whole package. On Windows,
Set-Content -Encoding UTF8adds one; use-Encoding utf8NoBOM(PowerShell 7+) or an editor that can save without one. The file looks correct either way, which is what makes this one nasty. Bothinstallandhoudini-packagewrite without a BOM.A second
fxhoudinimcp.json. Houdini processes every packages directory and the last one wins, so a leftover file can override a fresh install. Both commands warn when they find another one.fxhoudinimcp houdini-packagelists every one it can see, and what each points at, andfxhoudinimcp uninstallremoves the lot.No package file for the Houdini you launched. Each Houdini version reads its own preference directory, so a file in
houdini21.0/packagesdoes nothing for a Houdini 22 you start afterwards.installwrites to every candidate for exactly this reason; you only see this if you narrowed it with--houdini-dir, or if that Houdini'spackagesdirectory did not exist when you ran it. Create it and re-run.
On Windows, note that OneDrive's Documents redirection means a desktop-launched Houdini and a shell-launched one can resolve different preference directories. The package log is what settles which one your Houdini actually reads.
Checking what you are actually running
An editable install reports the version it was created at, not whatever the working tree has become since, so an old checkout can be running while the metadata claims otherwise:
python -m fxhoudinimcp --versionWorth checking first whenever a documented subcommand behaves as though it does
not exist. Before 2.5.0, an unrecognised argument was ignored and the MCP server
started instead, so python -m fxhoudinimcp install on an older install printed
a warning about not reaching Houdini and then sat there, looking like a hung
installer. It now exits with unknown command and the list of real ones.
Usage
Launch Houdini normally. The plugin auto-starts once when the UI is ready (controlled by FXHOUDINIMCP_AUTOSTART env var). The startup script uses uiready.py, which stacks correctly with other Houdini packages. You can also control it manually from the MCP menu (Start Server, Stop Server, Connect a Client, Server Status).
MCP > Connect a Client... prints the claude mcp add line for the port this
session actually ended up on, and copies it to the clipboard. That matters with
more than one Houdini open: a second session moves itself to the next free port,
so the configured port and the real one differ.
Startup verifies that Houdini's mcp.health endpoint answers from the current
Houdini process before printing that the server is ready. If your assistant
cannot reach Houdini after an app restart, call get_houdini_connection_status
for structured diagnostics, then relaunch Houdini or align FXHOUDINIMCP_PORT
and HOUDINI_PORT if another process owns the port.
Once connected, your AI assistant can:
"Create a procedural rock generator with mountain displacement"
"Set up a Pyro simulation with a sphere source"
"Build a USD scene with a camera, dome light, and ground plane"
"Create an HDA from the selected subnet"
"Debug why my scene has cooking errors"Environment Variables
Variable | Default | Description |
|
| Houdini host address |
|
| Houdini hwebserver port |
|
| Port for the Houdini plugin to listen on |
|
| Set to |
|
| Set to |
|
| Address the Houdini plugin binds. Loopback by default: the bridge runs arbitrary Python in your Houdini session and has no authentication, so only widen this on a network you trust |
|
| MCP transport ( |
|
| Logging level |
Development
# Install dev dependencies
pip install -e ".[dev]"
# Run linter
ruff check python/
# Run tests
pytest
# Run integration tests inside a real Houdini (requires a license seat;
# uses the newest installed Houdini, override with the HYTHON env var).
# Works on Windows, macOS, and Linux:
python tests/run_integration.py
# Convenience wrappers: tests/run_integration.ps1 / tests/run_integration.sh
# Contribute this machine's Houdini builds to the node-availability table and
# regenerate the version annotations in server_instructions.md:
python tools/gen_node_versions.py
# Regenerate the derived search hints and the plugin-command manifest
# (run gen_node_domains after gen_node_versions, it reads that table):
python tools/gen_node_domains.py
python tools/gen_required_commands.py
# Regenerate the node vocabulary tables in the workflow prompts. Edit the
# groupings in tools/prompt_vocab.json, never the tables in the markdown:
python tools/gen_prompt_vocab.py
python tools/gen_node_versions.py --check # verify the table against this machine
python tools/gen_prompt_vocab.py --check # needs no Houdini; runs in tests too
HYTHON=/path/to/hython python tools/gen_node_versions.py # one specific buildNo node name in prompts/markdown/ is hand-written any more. The tables come
from tools/prompt_vocab.json through gen_prompt_vocab.py, which rejects a
name no sampled build has and dates the ones that exist in only part of the
20.5-22.0 range. tests/test_prompt_vocab.py enforces both, and also checks the
hand-written prose around the tables, since that names nodes too, plus that every
shipped help page the prompts cite still resolves.
Prompt file layout
prompts/markdown/ has three subdirectories, so what a file is for is visible at
every call site (load_markdown("workflows/pyro.md")):
instructions/โ what the server tells every client at connect time.workflows/โ one guide per subject, named after the SideFX help scope it draws on, sopyro.mdpairs with thepyro/manual andsolaris.mdwithsolaris/. 31 of them.shared/โ fragments injected into the above (housekeeping.md,layout_on.md,layout_off.md), never served alone.
Most subjects are reached through houdini_workflow(topic), where topic is the
scope name, so adding a subject means adding a markdown file and nothing else. simulation_setup dispatches on its sim_type argument
through an alias map, because SideFX files FLIP under fluid/ and RBD under
destruction/ while users ask for "flip" and "rbd"; anything with no specific
guide falls back to dyno.md, the general dynamics one.
The server searches every help corpus the install ships, zipped or loose. On
a full 22.0 that is 56 scopes and 11,451 pages, including the workflow manuals
(pyro/, fluid/, vellum/, destruction/, model/, assets/, copy/) and
the unzipped ones (copernicus/, mpm/, heightfields/, ml/). It costs about
half a second of lazy load and ~69 MB inside Houdini, and nothing in the
assistant's context until a lookup happens.
tools/node_versions.json accumulates. It records which builds have been
sampled and what node types each had, so one installed Houdini is enough:
your build merges into the shared evidence and the annotations are derived from
everything sampled so far. A contributor with a single Houdini produces exactly
the same table as someone with six. If a version has never been sampled by
anyone, the generator says so rather than guessing, and --check reports only
contradictions with the builds you actually have.
That evidence file is ~1 MB and is not shipped. The generator also writes
python/fxhoudinimcp/data/sampled_versions.json, a few hundred bytes listing
only which versions have been sampled, which does ship: the server compares the
connected Houdini against it at startup and warns when a version has never been
checked, so a marker like (21.0+) silently covering a future 23.0 becomes
visible instead. get_houdini_connection_status reports the same thing. It is
advisory: build_network(dry_run=True) validates node types against the running
Houdini and cannot go stale.
If Red Giant / Maxon Universe is installed, its OpenFX plug-in crashes hou
initialisation on Houdini 20.5.487 and later, so hython cannot start at all.
Set HOUDINI_DISABLE_OPENFX_DEFAULT_PATH=1 when running any of the above.
This is a Houdini/Universe conflict, not something this repo causes.
Unit tests mock hou and run anywhere. The integration suite in
tests/integration/ executes all 188 commands against live Houdini via
hython โ including end-to-end user scenarios (procedural modeling,
simulation, animation, lookdev) โ and prints per-command timing and
coverage reports; it is skipped automatically when hou is not
available. tests/integration/perf_sweep.py benchmarks handlers on
large scenes, and python tests/integration/bridge_e2e.py validates the
full HTTP transport (real hwebserver in hython driven by the MCP
server's own bridge).
How It Works
Houdini Plugin (
houdini/): Runs inside Houdini's Python environment. Registers@hwebserver.apiFunctionendpoints that receive JSON commands. Useshdefereval.executeInMainThreadWithResult()to safely executehou.*calls on the main thread.MCP Server (
python/fxhoudinimcp/): A standalone Python process using FastMCP. Exposes 188 tools, 8 resources, and 9 prompts via the MCP protocol. Forwards tool calls to Houdini over HTTP.Bridge (
python/fxhoudinimcp/bridge.py): Async HTTP client that sends commands to Houdini's hwebserver and deserializes responses. Handles connection errors and timeouts.
What a call costs
That main-thread hop in step 1 is not free, and it is the single biggest thing
to know when driving this. hou.* can only run on Houdini's main thread, so
every command is queued with hdefereval and waits for the next event-loop
tick. Measured on Houdini 22.0.368 with an idle scene:
| 0.5 ms |
any real command, including | ~50 ms |
10 nodes created one call at a time | ~800 ms |
the same 10 nodes in a single round trip | ~66 ms |
The floor is flat: a trivial query costs the same as a real one, because you are
paying for the tick, not the work. So the cost of a session is set by how many
calls it makes, not how much they each do, and batching is worth roughly an
order of magnitude rather than being a matter of neatness. That is why the
server instructions tell an assistant to design a whole graph and submit it as
one build_network, and why set_parameters, connect_nodes_batch and
verify_network exist alongside their single-item equivalents.
Numbers are from one Windows machine and will move with hardware and with how busy Houdini is; the ratio is the durable part.
Contact
Project Link: fxhoudinimcp
License
MIT
This server cannot be installed
Maintenance
Resources
Unclaimed servers have limited discoverability.
Looking for Admin?
If you are the server author, to access and configure the admin panel.
Related MCP Servers
- Alicense-qualityDmaintenanceEnables natural language control of SideFX Houdini for tasks including node management, parameter editing, and geometry inspection. It leverages RPYC to execute Python scripts and manage scene data through an MCP-compatible interface.MIT
- Alicense-qualityAmaintenanceAn MCP server for controlling SideFX Houdini via hrpyc, enabling AI assistants to interact with Houdini sessions.46MIT
- AlicenseCqualityCmaintenanceConnects SideFX Houdini to Claude via the Model Context Protocol, enabling control of Houdini scenes, nodes, rendering, and more through 166 MCP tools.100MIT
- Alicense-qualityDmaintenanceEnables AI agents to directly control SideFX Houdini, including creating nodes, setting parameters, executing Python, capturing viewports, and rendering frames, via 57 MCP tools.2MIT
Related MCP Connectors
A comprehensive Model Context Protocol (MCP) server that enables AI assistants to control Unreal Eโฆ
Free public MCP for AI agents โ 193 tools, 44 workflows. No API key.
MCP server for Google Veo AI video generation
Latest Blog Posts
- Who's Calling? MCP Hosts Are an Identity Blind Spot (And the Spec Knows It)By Om-Shree-0709 on .mcpAgent IdentityOAuth 2.1
- Your AI Chatbot Just Exposed Your CEO's Salary to an InternBy Om-Shree-0709 on .Agent IdentityMCP SecurityOAuth Delegation
- Why MCP Servers Need Execution Sandboxing (And Why Your Current Stack Isn't Enough)By Om-Shree-0709 on .Agentic AiPrompt InjectionWebAssembly
MCP directory API
We provide all the information about MCP servers via our MCP API.
curl -X GET 'https://glama.ai/api/mcp/v1/servers/healkeiser/fxhoudinimcp'
If you have feedback or need assistance with the MCP directory API, please join our Discord server