AI Agent · SideFX Houdini

在 Houdini 里，
用自然语言建网。 Build node networks
with natural language.

一个基于 AI 的 Houdini 智能助手。自主 Agent 循环读取节点网络、创建与连接节点、编写 VEX、执行 Python，全程在一个迭代循环里完成。 An AI assistant for Houdini. An autonomous agent loop reads your node network, creates and wires nodes, writes VEX, runs Python — all in one iterative loop.

40+ 工具 · 六大 AI 提供商 · Plan 模式 · 大脑启发式长期记忆 · 插件系统 · 中英双语界面。无需 pip，依赖全部内置。 40+ tools · six AI providers · Plan mode · brain-inspired long-term memory · plugin system · bilingual UI. No pip — all dependencies bundled.

立即下载Download 安装指南Install guide v1.5.7
0+
内置工具Built-in tools
0
AI 提供商AI providers
0
运行模式Run modes
0
语言界面Languages
Houdini Agent
Provider
Model
/obj/geo1·2 nodes selected
给 grid 加 noise 让表面起伏，再 scatter 5000 个点。Add noise to a grid to displace it, then scatter 5000 points.
+3个节点已创建nodes created /obj/geo1/grid1 · mountain_vex · scatter1
VEX
// 抬升曲面，制造山脉起伏
float amp  = 2.5;
float freq = 0.35;
float n = noise(@P * freq);
@P.y += n * amp;
@Cd  = set(n, n*0.6, 1.0-n);

已搭好 grid → mountain_vex → scatter 链路。noise 通过 VEX 抬升 @P.y，scatter 输出 5000 点。需要我把点 copy 上实例几何吗？Built the grid → mountain_vex → scatter chain. Noise displaces @P.y via VEX; scatter outputs 5000 points. Want me to copy instances onto them?

◈Terrain Scatter Setup5 steps
1 Create grid (200 × 200)
grid1 · size 10 × 10 · rows/cols 200 · orient ZX
2 Displace with noise VEX
mountain_vex · run_over Points · amp 2.5 · freq 0.35
3 Scatter 5000 points
scatter1 · npts 5000 · relax 25 iters · seed 0
4 Align normals (polyframe)
polyframe1 · N from Tangent+Up · style first_edge
5 Copy instances to points
copytopoints1 · pack instance · inherit N / pscale
grid
noise
scatter
polyframe
Opus 4.8·12.4k / 200k ctx·$0.03
实时面板 · 试着输入并回车Live panel · type & press Enter
01 — Capabilities

不只是聊天，它会动手。Not just chat — it acts.

基于 OpenAI Function Calling 协议，Agent 自主规划、调用工具、检查结果、继续迭代，直到任务真正完成。 Built on the OpenAI Function Calling protocol, the agent plans, calls tools, inspects results, and keeps iterating until the task is truly done.

F.01

自主 Agent 循环Autonomous agent loop

接收请求 → 规划 → 调用工具 → 检查结果 → 继续，直到完成。多轮工具调用由 AI 自主决定顺序。Request to plan to tools to inspection to continue — multi-turn tool calls, ordered by the AI itself, until finished.

StreamingStoppableTodo system
F.02

VEX 与 Python 执行VEX & Python

优先创建 Wrangle 写 VEX；可在 Houdini Python Shell 运行 hou 代码，或执行系统 Shell 命令（带超时与安全保护）。Writes VEX in wrangles, runs hou code in Houdini's Python shell, or executes system shell commands — with timeouts and safety guards.

Wranglehou moduleShell
F.03

大脑启发式 长期记忆Brain-inspired memory

三层记忆系统 — 事件记忆、抽象知识、策略记忆，奖励驱动学习与自动反思。可全局开关。A three-tier system — episodic, abstract, and strategic memory — with reward-driven learning and auto-reflection. Globally toggleable.

EpisodicSemanticReflection
F.04

图片 / 多模态 输入Image / multimodal

为视觉模型附加图片，支持粘贴、拖拽、文件选择与缩略图预览。可截取 3D 视口直接喂给模型。Attach images for vision models — paste, drag-drop, file picker, thumbnails. Capture the 3D viewport straight into the model.

Paste / DropViewport capture
F.05

插件 与用户规则Plugins & rules

通过 plugins/ 目录扩展 Hook 事件、自定义工具与 UI。类似 Cursor Rules 的持久上下文规则，自动注入每次请求。Extend hooks, custom tools and UI via plugins/. Cursor-style persistent rules inject context into every request.

HooksCustom toolsRules
F.06

编辑式 深色界面Editorial dark UI

可点击节点路径、实时上下文栏、Token 分析、流式 VEX 预览、多会话标签页、中英双语动态重译。Clickable node paths, live context bar, token analytics, streaming VEX preview, multi-session tabs, dynamic bilingual retranslation.

Token analyticsMulti-session
Mode · 01

Agent

完整权限。可用全部 40+ 工具 — 创建、修改、连接、删除节点，设置参数，执行脚本，保存场景。Full access. All 40+ tools — create, edit, connect, delete nodes, set parameters, run scripts, save the scene.

Mode · 02

Ask

只读。仅查询场景结构、检查参数、搜索文档与分析；修改类工具被模式守卫拦截。Read-only. Query structure, inspect params, search docs, analyze — write tools are blocked by the mode guard.

Mode · 03

Plan

先调研、再澄清，生成带 DAG 流程图的结构化计划，用户确认后执行。内置自动续接防止过早终止。Research, clarify, then produce a structured plan with a DAG. Executes only after you confirm — with auto-continue against early stops.

Agent 循环The agent loop
用户请求Request 规划Plan 调用工具Call tools 检查结果Inspect 继续迭代Iterate 最终回复Answer
02 — Toolset

40+ 工具，一个注册中心。40+ tools, one registry.

统一的 ToolRegistry 集中管理核心工具、技能脚本与插件工具，支持基于模式的访问控制。A unified ToolRegistry manages core tools, skill scripts and plugin tools, with mode-based access control.

节点操作Node ops11
create_wrangle_node创建 Wrangle 并写入 VEX（优先）Create wrangle + VEX (preferred)
create_nodes_batch批量创建并自动连接Batch create & auto-connect
connect_nodes连接节点（可指定端口）Wire nodes by port
set_node_parameter设参 · 红绿 Diff · 一键撤销Set param · diff · undo
copy_node · delete_node复制 / 删除 · save_hip · undo_redoCopy / delete · save_hip · undo
查询与检查Query & inspect9
get_network_structure网络拓扑 · NetworkBox 感知Topology · NetworkBox-aware
semantic_search_nodes自然语言搜索节点类型Natural-language node search
get_node_parameters参数 + 状态 + 错误 + 连接Params + flags + errors + wiring
check_errors检查 cook 错误与警告Inspect cook errors / warnings
verify_and_summarize验证完整性并生成总结Verify & summarize network
代码 / 联网 / 文档Code · web · docs8
execute_python运行 hou 代码（30s 超时保护）Run hou code (30s guard)
execute_shell系统命令 · 安全检查 · 进程清理System cmds · safe · clean
web_search · fetch_webpage联网搜索与网页正文获取Search & fetch web pages
search_local_doc本地文档索引 · 节点帮助Local doc index · node help
capture_viewport视口截图喂给视觉模型Viewport capture for vision
布局 · 分组 · 性能Layout · group · perf10
layout_nodes自动 / 网格 / 分列布局Auto / grid / column layout
create_network_box语义颜色分组框Semantic-colored group boxes
perf_start_profileperfMon cook 时间 / 内存报告perfMon cook / memory report
create_plan · ask_questionPlan 模式：计划 + 澄清Plan mode: plan + clarify
run_skill · search_memory技能脚本 + 语义记忆检索Skill scripts + memory recall
03 — Models

接你想接的任何模型。Plug in any model.

云端、中转、本地、自定义端点 — 一个面板里随时切换提供商与模型。Cloud, relay, local, custom endpoints — switch provider and model from one panel.

DeepSeek
Reasoning
deepseek-v4-flashdeepseek-v4-pro
智谱 GLM
Stable in CN
glm-4.7
OpenAI
Function calling · Vision
gpt-5.2gpt-5.3-codex
Custom
任意 OpenAI 兼容端点Any OpenAI-compatible endpoint
LM StudiovLLMself-hosted
拼好饭 中转Relay
Claude · Gemini · MiniMax
claude-opus-4-6gemini-3.1-proMiniMax-M2.7
OpenRouter
16+ models · one key
claude-opus-4.6gpt-5.2grok-4.1-fast
05 — Install

三步开跑。Running in three steps.

无需 pip install — 所有依赖已内置在 lib/ 目录中。No pip install — all dependencies are bundled in lib/.

01

下载仓库Get the repo

Clone 或下载本仓库，放置到 Houdini 可访问的任意位置。Clone or download the repository to any location Houdini can reach.

02

在 Houdini 启动Launch in Houdini

在 Python 源码编辑器或 Shelf Tool 中运行启动代码。Run the launch snippet from the Python source editor or a shelf tool.

03

配置 API KeySet an API key

设置环境变量，或在工具内点击「设置 API Key…」保存到本机。Set an env var, or click "Set API Key…" in-app to save locally.

Houdini  20.5+ / 21+ Python  3.9+ (bundled) PySide  2 / 6 (bundled) OS  Windows · macOS
在 Houdini 中启动Launch in Houdini
Python
import sys
sys.path.insert(0, r"C:\path\to\Houdini-Agent")
import launcher
launcher.show_tool()
配置 API Key（环境变量，推荐）Configure API key (env vars, recommended)
PowerShell
# DeepSeek
[Environment]::SetEnvironmentVariable('DEEPSEEK_API_KEY', 'sk-xxx', 'User')

# OpenAI
[Environment]::SetEnvironmentVariable('OPENAI_API_KEY', 'sk-xxx', 'User')

# OpenRouter
[Environment]::SetEnvironmentVariable('OPENROUTER_API_KEY', 'sk-or-xxx', 'User')
06 — In practice

说人话，它来建。Say it plainly, it builds.

真实交互示例 — 从一句话到一整套节点网络。Real interactions — from a sentence to a full node network.

示例 · 散点流程Example · scatter
创建一个 box，在上面 scatter 500 个点，然后把小球复制到这些点上。Create a box, scatter 500 points on it, then copy a sphere onto the points.
✓create_nodes_batchbox → scatter → sphere
✓set_node_parameternpts = 500
✓connect_nodescopytopoints1
✓verify_and_summarize0.4s
完成。创建了 box1 → scatter1 → copytopoints1，球体模板半径 0.05，500 个点。Done. Built box1 → scatter1 → copytopoints1, template sphere radius 0.05, 500 points.
示例 · 编写 VEXExample · write VEX
给所有点添加随机颜色。Give every point a random color.
✓create_wrangle_nodepoint · Cd
VEX
@Cd = set(rand(@ptnum), rand(@ptnum*13.37), rand(@ptnum*7.13));
已创建 attribwrangle1，为所有点设置随机 Cd 属性。Created attribwrangle1, setting a random Cd on every point.
示例 · 分析属性Example · analyze
/obj/geo1/OUT 有哪些属性？What attributes does /obj/geo1/OUT have?
✓run_skillanalyze_geometry_attribs
该节点有 5 个 point 属性：P、N、Cd、pscale、id。5 point attributes: P, N, Cd, pscale, id.
示例 · 系统命令Example · shell
给 Houdini 的 Python 安装 numpy。Install numpy into Houdini's Python.
✓execute_shellhython -m pip install numpy
成功安装 numpy-1.26.4。Successfully installed numpy-1.26.4.
07 — FAQ

常见问题。Common questions.

+Agent 不调用工具怎么办？The agent isn't calling tools?
确认所选提供商支持 Function Calling。DeepSeek、GLM-4.7、OpenAI、拼好饭（Claude）、OpenRouter 均支持；自定义端点需指向支持工具调用的模型。Make sure the provider supports function calling. DeepSeek, GLM-4.7, OpenAI, the Claude relay and OpenRouter all do; a custom endpoint must point at a tool-calling model.
+节点操作失败？Node operations failing?
确认在 Houdini 内运行（非独立 Python），节点路径为绝对路径（如 /obj/geo1/box1），并查看工具结果中的具体错误信息。Run inside Houdini (not standalone Python), use absolute node paths like /obj/geo1/box1, and read the error in the tool result.
+API 连接有问题？API connection issues?
使用「测试连接」按钮诊断，检查 API Key 是否正确，并确认网络可访问 API 端点。Use the "Test connection" button, verify your API key, and confirm the endpoint is reachable.
+如何更新？会丢配置吗？How do updates work — do I lose config?
点击工具栏 Update 检查新版本；插件启动时静默检查 GitHub 并显示更新横幅。更新会保留 config/、cache/、trainData/、plugins/、rules/ 目录。Click Update in the toolbar; the plugin also checks GitHub at startup and shows a banner. Updates preserve your config/, cache/, trainData/, plugins/ and rules/ directories.
+数据隐私如何保障？What about privacy?
可通过自定义端点接入自托管模型（如 LM Studio、vLLM），全程离线推理，数据不出本机。云端提供商仅在你配置对应 Key 时调用。Point a custom endpoint at a self-hosted model (LM Studio, vLLM) for fully offline inference — nothing leaves your machine. Cloud providers are only called when you configure their keys.

把 AI 装进你的节点网络。Put AI in your node graph.

免费、开源。下载、解压、在 Houdini 里运行一行代码即可。Free and open source. Download, unzip, run one line in Houdini.

v1.5.7 · Houdini 20.5+ / 21+ · Windows & macOS · MIT
下载 ZIPDownload ZIP GitHub 仓库View on GitHub 历史版本Releases