Blog

/

Network rendering and AI agents on your farm

Network rendering and AI agents on your farm

Network rendering across your render nodes and workstations, with Claude Desktop or any other MCP client running renders and scripts on the farm.

Network rendering across your render nodes and workstations, with Claude Desktop or any other MCP client running renders and scripts on the farm.

RenderFlow

AI

Tutorial

The Claude crab throws a Blender file into a render farm: a server wired to six machines
Péter Sárhidai

Péter Sárhidai

14 min read


Network rendering used to mean one thing: send a scene to other computers and get frames back faster. That part still matters, and we’ll cover it properly. What’s new is who can use the farm. An AI agent connected to your farm can now open 300 scenes in 3ds Max, change the lights in Blender, build colour variations in Cinema 4D or resize 4,000 textures, spread over every free computer on your network. Then it renders every scene and passes the frames on to post-production: a Nuke comp, a video encode, each step starting when the one before it finishes.

This guide covers both halves. First what network rendering is and what an agent adds, then the practical part: hardware, Windows, Linux and macOS, network drives, ports and paths, for anything from a homelab to a small or mid-size studio. We’re 3D artists ourselves, and most of this we learned on our own projects.

Network rendering, the short version

A render farm manager sits between your artists and your machines. Someone submits a scene, the manager splits it into tasks (frames, chunks or tiles), hands each task to a machine that’s free and able to render it, and collects the results in one place. When a frame fails, it tries again, and the task log usually shows you why.

How network rendering works: submit a Blender job, split it into frame ranges, free machines render, the frames land in one folder

Figure 1. A hundred frames in ranges of ten. Each free machine takes the next range, and every frame lands in the same output folder.

Three things make it worth the setup:

  • Time. Ten machines render a 300-frame animation about 10 times faster than one.

  • Resources. Workstations sit idle from 7 pm to 9 am. A studio with 4 render nodes and 12 workstations has 16 machines for the night, without buying anything. When the machines you own aren’t enough, the same farm grows into your own Azure or AWS account.

  • Management. A job queue with priorities, dependencies and pools: urgent work goes first, a comp step waits for its render, and each job lands on machines that can run it. Failed frames retry on their own, and the artist who used to copy scenes around goes back to making them.

What about a single still? Split a high-resolution still into tiles: each machine renders a tile and RenderFlow puts the image back together. For 3ds Max with V-Ray or Corona, distributed rendering works too, with several machines rendering the same frame together. Either way, one big still finishes in a fraction of the time.

The tool that came with your application (Backburner, Team Render, Flamenco, HQueue) does this for one application. A farm manager such as RenderFlow does it for 26 applications on one farm, with Windows, Linux and Mac machines side by side, and automates and schedules the farm itself: idle workstations, power, pools and a weekly timetable.

What changed: the farm runs more than renders

A RenderFlow farm runs scripts as well as renders, and they’re split across machines the same way frames are. Script jobs run inside the application itself, without anyone opening it, such as MAXScript or Python inside 3ds Max, or Python inside Blender.

Each one takes a work list: a range, a count or a list of items, such as scene files, cameras, colourways or seeds. Every item becomes a task that any suitable machine picks up. Jobs can wait for each other, so a script that prepares 40 scenes can hand straight over to the render that follows.

That’s what makes the second half possible.

Your AI agent, with a farm behind it

Every RenderFlow server has a built-in MCP server. Connect Claude Desktop or any other MCP client and the agent can read the farm and act on it: submit renders and scripts, requeue frames, move machines between pools, look at rendered frames and set schedules.

Put that together with script jobs and an agent stops being a chat window. You describe the work, the agent writes the script, and the farm runs it in the right application on as many machines as fit.

Two examples, and what happens on the farm:

“Open every scene in this project, relink the missing textures to the new server, save them, and render a preview of each camera.”
The agent writes a MAXScript and submits it as a 3ds Max Script job with one task per scene. Every free machine opens a scene in 3ds Max in the background, relinks the textures, saves it and takes the next one. A render job waits behind the script and starts once the last scene is saved, with a preview of every camera. Forty scenes, twelve machines, one coffee.

“Set up ten camera angles around the headphones, light each one, and render.”
The agent writes a Python script for Blender and submits a Blender Script job with a work list of ten cameras, plus a render job that waits for it. Each task places its camera, aims a key light for that angle and saves its own .blend file. Free machines take one camera each, and once the last one is saved, the render starts and the ten shots land in the same output folder.

And a few casual ones for running the farm:

  • “How did the overnight renders go?”

  • “Which machines are idle right now?”

  • “Why did frame 137 fail?” The answer comes with the renderer’s own log lines and what your error catalog says about them.

  • “Requeue the frames that failed last night.”

  • “Are the last frames of my render actually there, or black?” The agent looks at the rendered previews.

  • “Move the workstations into the render pool for tonight.”

  • “Shut the render nodes down at 22:00 every night until Friday.”

How a script job runs: Claude sets up ten cameras in a Blender script, machines run it next to other work, the render follows

Figure 2. One task per camera. Machines set up the cameras in Blender in the background, and the render starts as soon as the last camera is saved.

You stay in charge

An agent with hands on your farm should have clear limits, and these are built in:

  • Off until you turn it on. AI clients stay off on a new farm until an admin switches them on under Settings → Security.

  • Nothing deletes. None of the MCP tools can delete a job, a machine or a file.

  • Check before it runs. Ask the agent to submit jobs suspended, look at the job and its script, then start it yourself.

  • Keys that do only what you allow. A laptop without RenderFlow connects with an API key, and a read-only key makes an agent that answers questions and changes nothing.

  • Your farm stays on your network. Clients connect to RenderFlow on the same computer, not through the internet. Want the model local too? LM Studio runs a local model that supports tool calling, so nothing leaves the building.

Claude answering questions about the farm with RenderFlow's tools: the overnight renders, a requeue, the idle machines and a nightly shutdown

Figure 3. Ask the farm in plain words: the agent checks it, requeues what failed and sets the schedule.

Render farm setup: what you need

Here’s the practical part, for a studio or a homelab.

The server

One machine coordinates the farm: it keeps the job queue and decides which machine renders what. That’s book-keeping, not heavy computing. We recommend 4 CPU cores, 16 GB of RAM and 100 GB or more of free space on an SSD.

A mini PC, an old desktop, a Linux box in the cupboard or a Mac mini will do. Give it a fixed IP address (a DHCP reservation on your router is enough), because all the other machines connect to it. And keep it out of the rendering: a heavy render fighting the database for memory slows the whole farm down.

Render nodes and workstations

Any machine that can run your render can join the farm:

  • Windows 10 or 11, or Windows Server 2019 and later

  • Linux: Rocky (8, 9, 10), Ubuntu and other RHEL or Debian based distributions

  • macOS 13 or later on Apple silicon

A few rules of thumb:

  • RAM matters most. Match your workstations. A scene that needs 90 GB on a workstation pages to disk or fails on a node with 64 GB.

  • Mixed machines are fine. RenderFlow benchmarks each machine and sends work to the fastest first, and a job can ask for a minimum of RAM or VRAM, so it never lands on a machine that’s too small for it.

  • Keep about 100 GB free on the system drive. Applications write their temporary files there while they render.

  • Install the same versions everywhere. The same application, renderer and plugins as the artist who made the scene. RenderFlow’s Inventory screen shows what’s installed on each machine and only sends a job to machines that have what it needs, but the versions still have to be there.

Mac mini as a render node? Yes, for anything that runs on macOS: Blender, Cinema 4D, Houdini, Maya, Nuke, After Effects, DaVinci Resolve, ComfyUI.

Shared storage

Every machine needs to open the same scene, find the same textures and write frames to the same place. That means a NAS or a file server, plus one shared folder RenderFlow uses for job files, logs and frame previews (the repository).

Farm size

Storage that keeps up

2 to 5 machines

A share on a fast SSD in one machine, or a small NAS

5 to 20 machines

A NAS with SSD cache, on wired Ethernet

20 and more

A file server or a NAS with 10 GbE, ideally SSD

Any NAS that shares over SMB or NFS works, in a homelab or a small studio.

Network

When ten machines open the same 4 GB scene at once, the network runs out long before the CPUs do. 1 GbE is workable for a small farm; 10 GbE is recommended and hard to avoid as the farm grows. Use cables, not Wi-Fi, for anything that renders.

Only the RenderFlow server needs internet access, for sign-in and licensing. Render nodes and workstations don’t need it for RenderFlow, though your applications and render engines may still need it for their own licences. For a farm with no internet at all, offline licensing (Enterprise only) runs the server from a licence file.

What a farm needs: a server in the middle, wired to render nodes and workstations above, your laptop with an AI client and a NAS below

Figure 4. One server in the middle. Render nodes, workstations, your laptop and the NAS connect to it, and only the server goes online.

Getting the paths right

Most first-night failures come from paths, so this gets its own section. A scene that renders fine on the artist’s machine finds its textures, caches and output folders through that machine’s own view of the network: a drive letter, a mount point, a path someone typed once. A render node sees the network differently, so the scene opens, can’t find its files, and renders black frames or fails outright. Get the paths right once and most of those problems never happen.

Use UNC paths, not drive letters

P:\projects\headphones\product_v03.blend works on the artist’s machine. On a render node running as a service, P: doesn’t exist: a drive letter belongs to a logged-in user, and a service has no user session. Point scenes and the repository at UNC paths instead:

\\nas\projects\headphones\product_v03.blend

If your studio lives on drive letters, add them under Settings → Mapped Drives, and each Windows machine will map them when RenderFlow starts.

Windows, Linux and Mac on one farm

The same share has a different path on each platform. Settings → Mapped Paths translates them once for the whole farm:

Platform

Path

Windows

\\nas\projects

Linux

/mnt/projects

macOS

/Volumes/projects

Mount the share in the same place on every Linux machine (an SMB or NFS entry in /etc/fstab) and every Mac, and a job submitted from Windows renders on all three.

Nothing on the desktop

C:\Users\peter\Desktop\textures renders on Peter’s machine and nowhere else. Keep every asset on the share. This one’s worth an email to the whole team.

Ports and firewalls

Port

Protocol

What for

44442

TCP

API and live updates (machines and agents connect here)

44443

UDP

Finding the server on the network

44444

TCP

The database, on the server itself

44445

TCP

The farm interface over HTTPS, when TLS is on

The installer opens these for you: Windows Firewall rules on Windows, firewalld or ufw on Linux, and Application Firewall exceptions on macOS. A third-party firewall needs them added by hand.

On a Mac, macOS asks the first time RenderFlow reaches another machine on the network. Press Allow for both RenderFlow and rfsv. If you missed it, turn both on under System Settings → Privacy & Security → Local Network.

On a homelab with VLANs, put the farm on one network, or type the server’s address on each machine if discovery doesn’t find it.

Setting up, step by step

  1. Install the server. Pick the server role in the setup wizard and sign in with your Pulze account. Only the server requires you to sign in. Render nodes and workstations join without an account of their own.

  2. Set the repository to a UNC path on your share, and test it from one other machine.

  3. Install the render nodes. Run them as a service (a Windows service, a systemd service on Linux or a launchd daemon on macOS) so they start on boot with nobody logged in. Press Discover or type the server’s address.

  4. Add the workstations. In Settings → Automation, turn on Use idle machines so a workstation joins the farm after its artist has been away for a set time, and Return to user so it goes back to the artist when they return.

  5. Render one real job. Render the project you’re working on right now, with its real textures and plugins, and check the frames.

  6. Connect your agent. Turn on AI clients under Settings → Security, then open Settings → AI Clients on the computer you work on and press Connect next to your client. For Claude Desktop, press Install, then Install again in Claude Desktop. There is nothing to fill in.

  7. Ask it something small first. “How is the farm doing?” is a good start, or “Requeue the failed frames of the last job.” Once that works, move on to scripts.

Deploying to 100 machines? The installer runs silently with flags for the role, the server address and the service, so a script or your deployment tool can set up every machine. The flags are listed in Silent Deployment.

The Machines screen: a Linux mini PC as the server, two Windows render nodes, a Mac mini and four workstations, each with its status and what it is working on

Figure 5. The Machines screen: a Linux mini PC runs the server, two Windows render nodes and a Mac mini take the work, and four workstations join when their artists are away.

Mistakes that cost a night of renders

  • Version mismatches. Someone updates Corona on their workstation and the nodes still run last month’s version. The frames render, but they don’t match. Check Inventory after every update.

  • Running out of licences. Ten nodes pick up a job when you own five render engine licences, and half the frames fail. Set a licence limit for the application, and the extra tasks wait for a free seat instead of failing.

  • Paths that only work on one machine. See above. Then see it again.

  • Failures nobody sees until morning. Turn on notifications for failed jobs (Slack, Teams, Discord, email or a webhook), or ask your agent first thing: “What failed overnight, and why?”

Try it on your own machines

Start with what you have: a server, a couple of nodes and a workstation are enough to see network rendering and an agent working on your own scenes. Try it on up to 10 machines for 30 days.

The quickest start: paste “Set up RenderFlow for me: https://renderflow.com/setup.md” into your agent. The agent walks you through installing the server and the nodes, and you can talk your network rendering setup through with it as you go.

Connecting your agent step by step is in the MCP server guide, and everything else RenderFlow does is on the RenderFlow page.

FAQ

What is network rendering?

Network rendering splits a render across several computers on your network. A render farm manager hands each frame or chunk to a free machine and collects the results, so ten machines finish a sequence about ten times faster than one.

Can an AI agent control my render farm?

Yes. RenderFlow has a built-in MCP server, so Claude Desktop or any other MCP client can submit renders and scripts, requeue frames, check results and manage machines. AI clients are off until an admin turns them on, and none of the tools deletes anything.

Can the agent work inside Blender, Cinema 4D or 3ds Max?

Yes, through script jobs. The agent writes a Python script (or MAXScript for 3ds Max), and RenderFlow runs it inside the application on the farm, split across machines: one task per scene, camera or variation.

Can a Mac mini be a render node?

Yes, on macOS 13 or later with Apple silicon, for any application that runs on macOS, such as Blender, Cinema 4D, Houdini, Maya and Nuke.

Can I run the server on Linux or a Mac?

Yes. The server runs on Windows, Linux and macOS. A mini PC is enough for a small farm.

Do I need 10 GbE?

Not to start. 1 GbE works for a small farm. As more machines open the same scenes at once, 10 GbE is the upgrade that keeps them fed.

Does my farm data leave the network?

No. RenderFlow runs on your own server and machines, and AI clients connect to it locally. Only the RenderFlow server talks to Pulze, for sign-in and licensing, and with offline licensing (Enterprise only) not even that.

More articles about

Render Flow

RenderFlow 2 is out: the Jobs view with a job rendering

RenderFlow 2, our render farm manager, is out: multi-step jobs, 26 applications, live render preview, artist-friendly logs and cloud machines in your own account.

RenderFlow

Release

Nov 10, 2026

4 min read
RenderFlow Updates - April 2026

RenderFlow in April 2026: cloud rendering for 3ds Max, Blender 5.0 and 5.1, FStorm 2, Octane 2026, new SDKs, Slack and Teams alerts, and new docs.

RenderFlow

Apr 2, 2026

2 min read
RenderFlow 1.1: New UI, User Groups, and Software Analytics

RenderFlow 1.1 brings a refreshed UI, granular access controls, user groups, software analytics, and community-requested features for studio teams.

Release

News

RenderFlow

Nov 25, 2025

3 min read

Built for freelancers, studios and creative teams

Take your workflow further

If this article was helpful, you’ll love what our tools can do inside your production pipeline.

Built for freelancers, studios and creative teams

Take your workflow further

If this article was helpful, you’ll love what our tools can do inside your production pipeline.

Built for freelancers, studios and creative teams

Take your workflow further

If this article was helpful, you’ll love what our tools can do inside your production pipeline.