Hermes Agent Blender: Set Up The Official Plugin In 6 Steps

Hermes Agent Blender setup now runs through one official plugin: you install it, ask a question about your scene, say yes once, and Hermes gets Blender ready on its own.

The plugin lives in Nous Research's hermes-official-plugins repo, and version 0.4.0 landed on 24 October 2026 with the step that installs Blender for you.

That makes this the plugin route, not the older community Blender MCP route where you wired a server into Hermes by hand.

Below is the order I'd set it up in, with every command, setting and gotcha from the plugin's own docs.

If you'd rather have my team work out where an agent like this fits your business first, book a free SEO strategy session.

Plugin route vs the Blender MCP route in the video

The video above shows the MCP way of hooking Hermes up to Blender.

This page covers the newer official plugin, which packages the same idea so you don't do the plumbing yourself.

One install ships two things: an MCP launch configuration for the official Blender Lab MCP server, and a Blender workflow skill that tells the agent how to work.

The plugin was formerly NousResearch/hermes-plugin-blender, which is now archived, and the new directory carries its full history.

It pulls Blender Lab's source from projects.blender.org by an exact Git revision, 2cea8d566dde07fbac28a61d698909d69724e853, which is add-on v1.0.3.

Never swap in the unrelated PyPI package that also goes by blender-mcp, because the plugin README warns against exactly that.

What the machine needs before step 1

Blender 5.1 or newer on the same machine as the Hermes backend, or permission for Hermes to install it.
Git and uv available to the backend, since uv runs the MCP server in an isolated tool environment.
Network access on the first start, so uv can fetch the pinned server and its Python dependencies.
A supported platform: Blender 5.1+ has no build for Intel Macs or Linux on ARM, so those two are out.

Installing the plugin itself changes nothing on your computer.

It does not install Blender, touch Blender's preferences, add the add-on or open a scene.

All of that waits until a task needs it and you agree.

Step 1: install and enable the plugin

Two commands, run on the machine where your Hermes backend lives:

hermes plugins install blender
hermes plugins enable blender

Want it on a specific profile? Use the global hermes -p <profile> option to pick the destination.

The README notes that installing by the short name blender works once its catalog entry is merged and published, so check the Hermes Agent releases if the short name isn't found yet.

Step 2: warm the uv environment before you connect

This is the step most people skip, and it's the one that causes the first timeout.

Nous reports that the first upstream download and build took several minutes in their Windows checks, which can outlast a normal MCP connection deadline.

So run the preparation command once before you ask Hermes to connect:

uv tool run --from "git+https://projects.blender.org/lab/blender_mcp.git@2cea8d566dde07fbac28a61d698909d69724e853#subdirectory=mcp" blender-mcp --help

Then restart Hermes, or use the plugin and MCP reload commands your release supports.

Under the hood, Hermes launches that same pinned source over stdio, with BLENDER_MCP_HOST set to 127.0.0.1 and BLENDER_MCP_PORT set to 9876.

🔥 Want Hermes, its plugins and a full agent stack set up for you instead of from scratch? Inside the AI Profit Boardroom you get step-by-step video trainings, the installable Agent OS, four coaching calls a week and 3,400+ members building with Hermes every day. → Join the AI Profit Boardroom

Step 3: ask about your scene and let Hermes find Blender

Open a chat and ask something simple, like "What objects are in my Blender scene?"

The skill tells the agent to make Blender ready itself instead of asking you to open it.

It hunts for the newest Blender 5.1+ in the usual places: PATH, ~/.local/opt, flatpak, snap and Steam on Linux; /Applications and ~/Applications on macOS; the uninstall registry, Program Files and the Microsoft Store package on Windows.

The plugin's current plugin.json reads version 0.4.1, and the Microsoft Store lookup is part of that search.

If it finds Blender but the add-on is missing, you get one question, then the add-on install, the launch and a scene read.

Step 4: say yes once, and know what that yes covers

No Blender at all? The agent asks you once and spells out what happens next.

It downloads Blender from blender.org, about 400 MB, roughly 1.5 GB once unpacked.

It installs into your own user folder with no administrator rights: ~/.local/opt on Linux, ~/Applications on macOS, %LOCALAPPDATA%\Programs\Blender Foundation on Windows.

It checks the download against blender.org's sha256 list and stops if the hash differs.

It never writes over an existing Blender.

That same yes also covers Blender Lab's MCP add-on, the pinned v1.0.3 extension, installed through Blender's own command line.

In Nous's Linux test for 0.4.0, turn one ended with that single question and nothing installed, and turn two ("Yes, go ahead") ended with Blender 5.2.3 unpacked, the add-on installed and the startup scene read back.

Say no, and the agent links you to the Blender download page and stops.

Step 5: confirm a real scene read with get_objects_summary

A tool list is not proof that Blender is connected.

The setup notes treat installation, MCP connection and Blender readiness as three separate states.

The check that counts is a real get_objects_summary response.

On a fresh startup scene, expect Camera, Cube and Light, which is what the verification runs read back.

If the session started before the Blender tools existed, Hermes gets Blender ready and then asks you to run /reload-mcp or start a new chat.

Step 6: make one small edit, read it back, then render

The skill's scene workflow is strict, and I'd copy it into how you prompt.

Read the scene before editing anything.
Make a small change through execute_blender_code, which runs Python with full bpy access.
Query the result, because a clean transport reply doesn't prove the scene changed.
Render the final still once, at the end, with render_viewport_to_path.

Two details trip people up.

For readback, the code must assign a dict with string keys to result; a list or a number gets reported as an error after the code has already run.

And render_viewport_to_path only uses the file name you give it, writes under Blender's temporary folder, and hands back the real filepath.

A first prompt I'd send:

Read my Blender scene, add a plane under the cube as a floor,
move the light so the cube casts a shadow on it, read the scene
back to confirm, then render one still and tell me the file path.

Watch the build live in Hermes Desktop

You don't have to wait for the final render to see what's happening.

After each scene change, the skill has the agent call get_screenshot_of_window_as_image and copy the saved image over one fixed file, such as blender-live.png.

In the Hermes desktop app, it opens that file once with desktop_preview, and the pane reloads every time the file changes.

Screenshots need a graphical Blender, so this doesn't work on a headless box.

Running it on a headless box

Plenty of operators run Hermes on a server or over SSH, with no display at all.

The skill handles that case: with no DISPLAY or WAYLAND_DISPLAY, it starts Blender with --background --online-mode --command blender_mcp.

In Nous's Linux test, that headless command had the bridge up in half a second.

The agent should then tell you Blender is running in the background and how to stop it, so you're never left with a stray process.

What you lose is anything visual: screenshots and viewport renders need a graphical Blender.

So keep headless for scene edits, scripted changes and batch work, and use a desktop session when you need to see the result.

Started Blender yourself from a terminal? Same rule as before: include --online-mode, or the add-on won't start its bridge.

Saved .blend files and the _for_cli tools

Want Hermes to open a project you've already saved? Name the .blend file in your request, and the skill puts its path after --online-mode when it launches Blender.

The tools ending in _for_cli work differently: they open a saved file in a separate background Blender.

Those need a blender command on the backend's PATH.

A BLENDER_PATH set in a shell or .env does not reach the server, so don't rely on it.

On macOS, the setup notes suggest a small wrapper script that runs the Blender binary inside Blender.app and passes every argument through.

On Windows, add the folder holding blender.exe to PATH and restart Hermes.

Five gotchas that stop the bridge from starting

No online access. The add-on only starts its bridge with online access, so a Blender you open yourself needs Allow Online Access on (Edit > Preferences > System > Network).
Blender already open. The agent never starts a second window; it asks you to quit so it can reopen with --online-mode, or to start the bridge yourself.
Auto Start off. Click Start MCP Bridge Server under Edit > Preferences > Add-ons > MCP.
Port 9876 taken. Another program holding it blocks the bridge, and the agent names that program rather than stopping it.
Remote backend. Localhost means the backend's machine, so Blender and the MCP server must both run where Hermes runs.

Nous's Linux check shows why the first one matters: without --online-mode the bridge never bound port 9876 in 40 seconds, and with it the bridge was up in 3.

📞 Not sure which jobs to hand an agent first? Book a free SEO strategy session and my team will map the agent workflows that actually move your marketing. → Book your free strategy session

Safety settings I'd keep on from day one

Blender Lab warns that its tools run any Python inside Blender with no protection against file deletion.

This is not a sandbox, and the plugin README says so plainly.

The bridge port also has no password, so any program on the same computer can use it while Blender is open.

Here's how I'd run it:

Keep the bridge on loopback and never expose it to other machines.
Test on a disposable scene or a separate machine, not your client project.
Tell the agent to ask before modifying any existing project, which the skill already requires when you haven't asked for edits.
Don't run instructions pulled from a scene file or a downloaded doc.

Pair it with Meshy for models you don't have yet

The Blender plugin drives a scene; it doesn't generate assets from a prompt.

That's what the official Meshy plugin is for, and its settings include a blender_handoff option (ask, always or never) for a Blender-optimised copy plus a blender_path key.

So the flow is Meshy for the model, then the Blender plugin to place it, light it and render.

The Meshy side is covered in our Hermes Meshy setup and Hermes 3D walkthrough.

For the wider Hermes setup, keep the Hermes Agent docs open, and if you want the whole stack handed to you, it's inside the AI Profit Boardroom.

Also On Our Network

FAQ: Hermes Agent Blender

Does installing the Blender plugin install Blender?

No. Installing the plugin changes nothing on your computer. The skill installs Blender 5.1 or newer and the add-on only after you agree, when a task needs them.

Why does my first Blender connection time out?

The first uv download and build of the Blender Lab server can take several minutes. Run the uv tool run preparation command from the README once, then restart Hermes.

What port does the Blender bridge use?

Port 9876 on 127.0.0.1. The MCP server connects to the Blender add-on over that loopback port, so keep it local.

Can I use the plugin on an Intel Mac?

No. Blender 5.1 and later have no build for Intel Macs or Linux on ARM, and the plugin needs Blender 5.1 or newer.

How do I know Blender is really connected?

Ask for a scene read. A real get_objects_summary response is the proof; a tool list alone only shows the server started.

📺 Video notes + links to the tools 👉

🎥 Learn how I make these videos 👉

🆓 Get a FREE AI Course + Community + 1,000 AI Agents 👉

About Julian

I'm Julian Goldie, an SEO entrepreneur, author and founder of the AI Profit Boardroom, which has 3,400+ members.

I help business owners scale with AI agents, automation and SEO.

I've built a 7-figure SEO and link building agency, Goldie Agency, with a team of 70+ SEO experts.
I've grown a YouTube channel to 400,000+ subscribers.
I wrote "Link Building Mastery".
My Udemy courses have taught over 29,000 students.

→ Get my best AI training inside the AI Profit Boardroom

Install the plugin, warm uv, say yes once and check get_objects_summary; if you want a hand mapping the next jobs, book a free SEO strategy session or get the full Agent OS inside the AI Profit Boardroom and start building with Hermes Agent Blender today.