Awesome-Gaussian-Skills: Connect 1,000+ 3DGS Papers and 34 Rendering/Modification Tools to Your AI Agent

jaccen/Awesome-Gaussian-Skills is an AI-native entry point for 3D Gaussian Splatting (3DGS) research: it bundles 1,000+ arXiv-verified research methods, 15 AI agent skills, and an MCP server with 34 tools. It is best suited to the task of letting an AI agent read and retrieve 3DGS research knowledge while directly driving a gsplat rendering loop. The most important prerequisite before use: the MCP server is v0.8.0 and requires a local Node runtime; the 13 experimental tools are only stubs, and what you can actually rely on are the 21 core tools.

Who It's For

• People doing or studying 3DGS research who need to quickly browse 26 categories and 1,011 research entries (snapshot dated 2026-10-21) and would like an agent to retrieve and summarize them on their behalf — 904 of these entries already have freshly verified arXiv metadata.
• Users who want to add 3DGS domain knowledge to an existing agent (skill-catalog-style integration, such as a skills directory like .claude/): each skill is a standalone SKILL.md file that works once copied into the skills directory.
• Developers who want to manipulate 3D scenes, cameras, and Gaussians and render frames through MCP tools.
• When it's not a fit: if you need a runtime other than Node, the evidence does not say whether that is supported; users depending on experimental features should note that these 13 experimental tools are stubs; 107 research entries have not yet completed first-source verification, so check the original papers before citing them. The license is Apache-2.0.

First Attempt in 3 Minutes

The most direct main path given officially is copying the skill files plus configuring the MCP server over stdio, in a few steps:

1. Prerequisites: Node is installed locally (the MCP server starts with command: node); clone this repository.
2. Hook up the skills: per the repository README, "each skill is a standalone SKILL.md file — copy it into your agent's skills directory." Note: copying the skills content into your agent's skills directory (the example directory given in the README is .claude/) is a write to your local agent configuration directory, not a read-only operation; the official evidence provides no restore or rollback instructions, so whether and how to undo it is up to you to decide and implement.
3. Configure the MCP server: in your MCP client configuration, use the officially provided stdio configuration snippet — command: node, args: [/path/to/mcp-server/dist/index.js], with stdio as the transport. Replace /path/to/... with your local repository path.
4. First low-risk call: start by calling query and browse capabilities to verify connectivity, rather than directly calling tools that rewrite the scene. Among the core tools, render_frame and set_camera operate on the local rendering session and are lower risk than rewriting Gaussian data; if you want to call write operations such as modify_gaussians or sculpt_pipeline, note that they modify the Gaussians in the session, the official evidence describes no undo path, so run them on worthless test data.
5. Success signal: if the client can list the server's 34 tools (21 core + 13 experimental), the stdio connection is working; the repository also states that the MCP server exposes the gsplat rendering loop over HTTP + WebSocket on port 9842, which you can use to further confirm the rendering service is running.

Note: tools being listed and callable only shows that the connection and session are usable; it does not mean the rendering results are correct or the data is intact. Visually confirm the output of the first rendered frame yourself.

Use Cases

Case 1: Retrieving a 3DGS research direction. Scenario: you want an agent to summarize a class of 3DGS methods. How to try it: after hooking up the 15 skills, have the agent retrieve and summarize based on the repository's registered snapshot (1,011 entries, 26 categories). What to expect: the agent can cite entries with verified arXiv metadata; you will see some entries flagged as still needing first-source verification (107 entries), and those need manual checking.

Case 2: Driving a local gsplat rendering loop. Scenario: you have a 3D scene and want an agent to adjust the camera and produce a rendered frame. How to try it: after starting this MCP server, call the core tools in sequence — import_scene to import the scene, set_camera to set the camera pose, render_frame to render a frame. What to expect: the server provides the gsplat rendering loop on port 9842 (HTTP + WebSocket), and you can observe frame output in the rendering session. The evidence does not provide the specific parameter schemas for each tool, so before the first call, check the tool descriptions returned by your client to confirm the parameters.

Boundaries of Use

• Read/write nature: copying the skills is a write to your local agent configuration directory; on the MCP server side, import_scene, modify_gaussians, set_camera, render_frame, and sculpt_pipeline operate on local rendering session data, and the tools that modify Gaussians actually change session state; the official evidence describes no undo mechanism.
• Runtime conditions: the v0.8.0 MCP server requires Node; the HTTP/WS port is fixed at 9842, and if there is a port conflict you must handle it yourself (the official documentation gives no instructions for changing the port).
• Data limitations: the knowledge base snapshot is dated 2026-10-21, and 107 entries are pending first-source verification — check the original arXiv papers before citing conclusions.
• Unconfirmed information: the parameter schemas of each tool, authentication methods, and multi-client compatibility are not covered in the evidence; before use, rely on the tool descriptions actually returned by your client.
• This write-up does not constitute a security conclusion: "operating a rendering session" is not the same as zero risk; boundaries such as network behavior and the scope of local data reads are not covered by the evidence and must not be assumed.