Provides tools for interacting with SideFX Houdini through a Houdini bridge, enabling agents to inspect scene information, list/create/delete nodes, set parameters, connect nodes, execute Python, save/load HIP files, and list node types.
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., "@Houdini MCP create a box node and connect it to a transform node"
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.
OpenAI-compatible Model Context Protocol server for SideFX Houdini.
Repository: github.com/alexantoshuk/houdini-mcp
Works with any agent path available in Zed:
Agent path | How it gets the tools |
Zed Agent | Uses Zed |
External Agents (Cursor, Claude ACP, Goose, Cline, OpenCode, Qwen Code, …) | Zed forwards configured MCP servers over ACP |
Terminal / other MCP clients |
|
Zed agents --stdio MCP--> houdini-mcp (Python)
|
+-- mock backend (tests / offline)
+-- TCP JSON-lines --> Houdini bridge (hou.*)Related MCP server: Houdini MCP Server
pip install -e .
python -m houdini_mcp --mock # offline smoke.\scripts\install_zed.ps1Then in Zed: Extensions → Install Dev Extension → select this repo folder.
Or add manually to %APPDATA%\Zed\settings.json:
"context_servers": {
"houdini-mcp": {
"command": "C:/Python310/python.exe",
"args": ["-m", "houdini_mcp"],
"env": {
"HOUDINI_MCP_HOST": "127.0.0.1",
"HOUDINI_MCP_PORT": "18812"
}
}
}Green status under Settings → AI → MCP Servers means the server is active for all Zed agents.
.\scripts\install_houdini_package.ps1Restart Houdini. The bridge autostarts on 127.0.0.1:18812 via scripts/123.py / 456.py — GUI sessions only. hython / hbatch skip the bind so they no longer steal the port from a live Houdini (no more renaming hooks to .bak).
Manual start / recovery inside Houdini Python Shell:
import houdini_mcp_bridge
houdini_mcp_bridge.ensure_bridge() # preferred — no-op if already listening
# or: houdini_mcp_bridge.start_server(port=18812)Env | Effect |
| Never autostart |
| Force autostart even in hython |
Focused tools with input schemas + safety annotations (readOnlyHint / destructiveHint):
get_scene_info, list_nodes, get_node
create_node, set_parameter, connect_nodes, delete_node
execute_python, save_hip, load_hip, list_node_types
OpenAI company-knowledge pair: search, fetch
Workflow skills: list_skills, get_skill — load SKILL.md from the public repo
alexantoshuk/houdini-mcp
(raw.githubusercontent.com / Contents API, ref main). Local .cursor/skills/ is offline fallback only.
Also exposes resources houdini://scene/info, houdini://skills/index, and prompt houdini_starter_network.
Env | Meaning |
| GitHub |
| Branch/tag/commit (default |
|
|
MCP instructions tell the agent to call get_skill(...) when the task matches. Push skill edits to main before expecting remote agents to see them.
Transport | Command | Use |
stdio (default) |
| Zed, Claude Desktop, Cursor, Codex |
Streamable HTTP |
| OpenAI remote MCP / Inspector |
pip install -e ".[dev]"
pytest -qProject skills live in .cursor/skills/houdini-*/SKILL.md (33 skills):
Solaris / assets / model / VEX — houdini-solaris, houdini-assets, houdini-model, houdini-vex
Sims — pyro, fluid, vellum, destruction, mpm, ocean, grains, …
Pipeline — tops, io, render, shade, character, crowds, …
Defaults — houdini-discipline (senior operating rules)
Index + attribution: skills/README.md.
Upstream guides (MIT): skills/vendor/fxhoudinimcp/.
Variable | Meaning |
| Bridge host (default |
| Bridge port (default |
|
|
|
|
|
|
|
|
Wire an output of from_path into an input of to_path.
Inputs: to_path (required), from_path (required), to_index, from_index. Returns: object
Create a Houdini node under parent. Example: parent=/obj, type=geo, name=geo1.
Inputs: type (required), parent (required), name. Returns: object
Delete a node and its descendants. Refuses to delete /obj.
Inputs: path (required). Returns: object
Run Python in the Houdini process with hou available. Prefer typed tools when possible.
Inputs: code (required). Returns: object
OpenAI-compatible fetch of a documentation item by id from search results.
Inputs: id (required). Returns: object
Fetch details for one node: type, parms, inputs, outputs, children.
Inputs: path (required). Returns: object
Use when you need Houdini session status: hip path, version, frame, fps.
Inputs: no named parameters. Returns: object
Load a houdini-* SKILL.md by name (e.g. houdini-camera, houdini-solaris). Call this when the user task matches that skill's topic — especially from workspaces that are not the houdini_mcp repo — then follow the skill text.
Inputs: name (required). Returns: object
List child nodes under a parent path. Set recursive=true to walk descendants.
Inputs: path, recursive. Returns: object
List available Houdini node types, optionally filtered by category (Object, Sop, Dop, ...).
Inputs: category. Returns: object
List houdini-* Agent Skills shipped with this package. Use when choosing which get_skill to load for the current Houdini task.
Inputs: no named parameters. Returns: object
Load a hip file into the current Houdini session.
Inputs: path (required). Returns: object
Save the current hip file. Provide path to save-as; omit to save current file.
Inputs: path. Returns: object
OpenAI-compatible search over node-type documentation. Returns id/title/url results.
Inputs: query (required). Returns: object
Set a parameter on a node. Value may be number, string, or list for multi-component parms.
Inputs: parm (required), path (required), value (required). Returns: object
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 15 tools
Most tools map cleanly to a distinct resource+action (get_node vs list_nodes vs create_node vs delete_node), and the skill tools are clearly separated. Minor ambiguity: generic `search`/`fetch` names sit alongside Houdini-specific node/doc tools, and `execute_python` can subsume nearly every other tool, though the description frames it as a fallback.
Strong, consistent snake_case verb_noun pattern across the board (get_scene_info, list_nodes, create_node, set_parameter, save_hip, get_skill). The only deviations are `search` and `fetch`, which are bare verbs retained for OpenAI compatibility, but they remain readable and non-conflicting.
15 tools is well-scoped for a DCC-integration server: a core scene-graph surface plus file ops, docs, and skills. Each tool earns its place and the set avoids both thin single-tool surfaces and sprawling 40+ tool menus.
Core lifecycle is covered: create/get/list/delete nodes, set parameters, wire connections, plus hip save/load and scene introspection. Gaps like disconnect, duplicate/rename, node flags (display/render), or geometry readback exist, but `execute_python` provides a workable escape hatch.
Host any HTTP API as a standard MCP server; auto-distribute to 12+ mainstream agent platforms.
Nifty's MCP server — exposes tasks, projects, messages, and files as tools for AI agents.
Remote MCP server exposing SMI Aware tools, resources, and skills over Streamable HTTP.
Remote MCP server exposing your Node Command fleet (files, shell, processes, browser) to AI hosts.
