MeshGen: An AI agent panel inside Blender, with five ways to point it at a model
Use AI Agents directly in Blender.
At a glance
- What is it?
- MeshGen is a Blender add-on that turns natural language prompts into scene edits through a configurable LLM backend. The interesting part is the backend matrix, and the part to check before adopting is the hardware floor it quietly imposes.
- Who is it for?
- Adopt MeshGen if you already run Blender on a machine with an NVIDIA GPU of at least 8GB VRAM, or if you are comfortable routing prompts to a hosted API and paying per call. Skip it if you work in a locked-down studio where scene data cannot leave the workstation, since the recommended path sends prompts to a remote provider.
- Can I use it commercially?
- Yes. MIT is a permissive licence: you can use, modify and sell software built on it, as long as you keep its copyright and licence notices.
- Is it still maintained?
- Yes. The repository last received commits 113 days ago.
- What is it written in?
- Mainly Python, according to GitHub's language statistics.
Answers come from the project's GitHub data, last synced on September 15, 2026, and from our analysis. They are not legal advice.
DEEP OPEN-SOURCE ANALYSIS
The gap MeshGen fills between a chat window and the viewport
Most attempts to put a language model near a 3D tool end in one of two places: a chat interface that produces code you paste into a script editor, or a generator that returns a finished asset with no scene context. MeshGen takes a third position. It installs as a Blender add-on, exposes a sidebar panel under the N key, and lets the agent act on the scene you already have open. The README frames the intent as AI as a tool, not replacement, and the interaction model matches that: you type a prompt such as Create a snowman, press Submit, and the agent works inside your file rather than handing you a download.
The audience is narrower than that phrasing suggests. The add-on is aimed at people who are already Blender users and who want to delegate repetitive or exploratory modelling steps without leaving the application. It is not a pipeline component, there is no CLI, and nothing in the README suggests a headless or batch mode. If your workflow is scripted asset generation running on a render farm, this is not the shape of tool you are looking for.
Five backends, two very different cost and privacy profiles
The design decision that defines MeshGen is that it does not ship a model. It ships a provider abstraction, and the provider you pick changes almost everything about how the add-on behaves.
On the local side there are two routes. The built-in Local backend runs inference directly inside Blender, and the README is explicit about the prerequisites: a powerful NVIDIA GPU with at least 8GB of VRAM, a cuda build of the addon, and a preference for in-process inference over a separate Ollama server. Setup is a button labelled Download Recommended Model, which fetches Meta-Llama-3.1-8B-Instruct-GGUF, or a manual drop of any .GGUF file into the models folder, reachable through a folder icon in the preferences panel. The second local route is Ollama: install Ollama, run ollama serve in a terminal, select Ollama in the Provider dropdown, and enter the server endpoint and model name. The README notes the defaults should work for most users.
On the remote side there are three providers. Hugging Face is the recommended default and the README describes it as providing limited free use of powerful models; you create a token at hf.co/settings/tokens and paste it into the API Key field, with an optional Model ID such as meta-llama/Llama-3.3-70B-Instruct. Anthropic and OpenAI follow the same shape, with their own key pages and example model IDs (claude-3-5-sonnet-latest and gpt-4o-mini respectively).
The practical split is this: local backends keep your scene on your machine and cost nothing per prompt, but demand specific hardware. Remote backends remove the hardware constraint and put your prompt, and whatever scene context the agent sends with it, on someone else's servers. The README does not describe what portion of the scene is transmitted, which is the single most important thing to determine before using a remote provider on client work.
Installing the add-on and picking a backend
Installation is manual. Download the addon ZIP from the latest release page, then in Blender go to Edit, Preferences, Add-ons, click the top-right arrow, and choose Install from Disk. Select the ZIP. After that, the add-on appears in the same Add-ons list under meshgen, and its preferences pane is where everything else happens.
That pane is where you select Local or Remote. Choosing Remote reveals the Provider dropdown with the Ollama, Hugging Face, Anthropic and OpenAI entries described above, plus API Key and Model ID fields. Choosing Local reveals the model download button and the models folder shortcut.
Usage after setup is deliberately minimal: press N to open the sidebar, select the MeshGen tab (or go through View, Sidebar), type a prompt, click Submit. There is no command palette, no keybinding list, and no scripting API documented in the README. If something goes wrong, the troubleshooting section points at the system console: on Windows, Window then Toggle System Console; on macOS and Linux, launch Blender from a terminal so stdout is visible. That is the whole debugging story as documented. There is no log file mentioned, so errors that occur before the console is open are gone.
LLaMA-Mesh and Hyper3D, and why they are not both local
Two optional integrations live under Edit, Preferences, Add-ons, meshgen, Integrations. The README states that when they are enabled, the agent is automatically given access to those tools depending on context, which means the agent's available action set is not fixed. It expands and contracts with your configuration. That is worth knowing when a prompt behaves differently on two machines.
LLaMA-Mesh provides local mesh understanding and generation, built on the nv-tlabs/LLaMA-Mesh project. Its requirements are the same GPU floor as the local backend: an NVIDIA GPU with at least 8GB VRAM and a cuda build of the addon. The third requirement is the counterintuitive one. You must be using a remote API backend, because LLaMA-Mesh loads locally on your machine while the agent reasoning runs remotely. So the combination is a local mesh model paired with a hosted language model, not an all-local setup. Enabling it means clicking Load LLama-Mesh and waiting.
Hyper3D is a hosted service at hyper3d.ai for high-fidelity mesh generation. You check Enable Hyper3D and enter an API key; the README notes that free use is currently provided with the awesomemcp key. It also warns that generation may take several minutes per mesh. Several minutes is a long time to hold a Blender session open, and the README gives no progress indicator or cancellation behaviour. Treat Hyper3D as a background task you start and then leave, not something you iterate on interactively.
Where MeshGen is the wrong tool
The hardware requirement is the first filter. If you do not have an NVIDIA GPU with 8GB or more of VRAM, the local backend and LLaMA-Mesh are both closed to you, and the README's wording (only select this option if) is a warning rather than a suggestion. AMD and Apple Silicon users are not addressed anywhere in the material. That is a real exclusion, not a documentation gap.
The second limitation is confidentiality. The README recommends Hugging Face for most users. That recommendation means most users are sending prompts to a third party. For a hobbyist building a snowman, fine. For a studio under NDA, the recommendation is unusable without first establishing what the agent transmits, and the README does not answer that question.
The third is reproducibility. The agent's tool access changes based on which integrations you enable, and the model behind it changes based on the Model ID you type into a free-text field. Two artists on the same project with the same add-on version can get different results from the same prompt. Nothing in the README describes a way to pin or record the configuration alongside the output.
Finally, there is no headless mode. Everything routes through the sidebar panel. If you wanted to script MeshGen into a build step, the material gives you nothing to work with.
How this compares to driving Blender from a coding assistant
The obvious alternative is not another Blender add-on. It is using a general coding assistant to write bpy Python that you run through Blender's script editor or the --python command-line flag. That approach has different properties in every dimension that matters here.
It requires no GPU, because the model runs wherever your assistant runs. It requires no add-on installation and no cuda-specific build. It produces a text artifact you can commit, review, and rerun, which addresses the reproducibility problem above. And it works headlessly, so it fits a pipeline.
What it does not do is see your scene. A coding assistant writes bpy calls against an imagined scene graph; MeshGen's agent operates on the file you have open, which is the entire premise of the add-on. The trade is context for auditability. If your task is a one-off modelling step where describing the scene in words is faster than inspecting it in code, MeshGen wins. If your task is a repeatable transformation you will run fifty times, generated bpy scripts are the better fit, and the fact that MeshGen has no documented way to export or replay an agent session is the reason.
Maintenance, licensing, and what the release cadence tells you
MeshGen is MIT licensed, which is permissive: you can use, modify and redistribute it, including in commercial work, provided the licence notice is preserved. That covers the add-on itself. It does not cover the models you point it at. Meta-Llama-3.1-8B-Instruct, the model behind the Download Recommended Model button, carries its own licence and acceptable use policy from Meta, and the same applies to whatever Model ID you type into the Hugging Face, Anthropic or OpenAI fields. Hyper3D is a separate hosted service with its own terms. The MIT grant on the repository tells you nothing about any of those. This is not legal advice; check each model and service licence against your use case.
On maintenance, the release history shows v0.6.0 in March 2025, v0.7.0 in April, and v0.7.1 two weeks later, with the last push to the default branch in May 2026. A patch release two weeks after a minor version suggests active bug-fixing rather than a stalled project, and the gap between the April 2025 release and the May 2026 push is not explained by anything in the supplied material, so treat the current activity level as unverified.
Upgrade cost is low by construction. Installation is a ZIP you replace through Install from Disk, and configuration lives in Blender preferences rather than project files. That also means your API keys and provider settings are stored in Blender's preference store, not in your .blend, so they do not travel with the file and do not appear in version control. Set them again on each machine.
Editorial conclusion
Adopt MeshGen if you already run Blender on a machine with an NVIDIA GPU of at least 8GB VRAM, or if you are comfortable routing prompts to a hosted API and paying per call. Skip it if you work in a locked-down studio where scene data cannot leave the workstation, since the recommended path sends prompts to a remote provider. Before installing, verify which build you are downloading: the local backend and LLaMA-Mesh both require the cuda variant of the addon, and a non-cuda ZIP will not give you either.
Community notes