Add Arynwood MCP, a free local-first AI workspace, to your RackWorks desktop VPS. Chat, knowledge base and design canvas, on the CPU.

Esta guía solo está disponible en inglés por el momento.

Arynwood MCP is a local-first AI desktop workspace for Linux, made by Arynwood Technology. It is free and open source (source on GitHub) and currently in alpha, so expect rough edges. It combines chat with Ollama and multiple personas, knowledge base search, a design canvas, and MCP tool integrations. RackWorks installs version 0.4.6 as a one-click marketplace app on top of the desktop on your VPS.

Read this first: what you get on a VPS

A RackWorks VPS has no GPU. On a VPS, Arynwood MCP gives you chat, knowledge base search, and the design canvas, running on the CPU. Because every response is computed on the CPU, replies are slow, and they get slower with larger models. The first reply after a quiet period is the slowest, because the model has to load into memory.

The features that need an NVIDIA GPU will not work on a VPS: Stable Diffusion image generation, text to speech, voice conversion, video, and LoRA training. They may appear in the app, but they cannot run here.

Requirements

RequirementDetails
Plan typeKVM VPS (not LXC)
Operating systemUbuntu 22.04, 24.04 or 25.04, or Debian 12 or 13
DesktopA desktop environment must already be installed (GNOME, KDE Plasma, XFCE or Cinnamon). See How to Install a Desktop Environment on Your VPS.
CPUAt least 4 vCPU
MemoryAt least 8 GB RAM
StorageAt least 30 GB, which covers the desktop plus about 8 GB for this app and its models
Install timeAbout 20 minutes (an estimate, mostly the model downloads)

The 4 vCPU and 8 GB figures are both the recommendation and the enforced minimum. If your plan is smaller, the app shows as unavailable on the Apps tab and tells you what the plan is short of. An upgrade clears it.

Installing Arynwood MCP

  1. Install a desktop environment first, if your VPS does not have one.
  2. Log into the Client Portal and select your VPS under "My Services."
  3. Click the Apps tab, find Arynwood MCP in the AI Tools group (the portal may show the group as just "AI"), and click Install.
  4. If no desktop is installed, the install is refused with a message saying so. Nothing is changed on your server.
  5. The card shows progress: Queued, then Installing, then Installed.

The installer sets up everything the app needs, and you do not run any commands yourself:

  • The Arynwood MCP app. The download is checked against its published SHA-256 checksum, and the install stops if it does not match.
  • Ollama, installed with its official installer.
  • Qdrant, the vector database behind knowledge base search.
  • The nomic-embed-text embedding model and the hermes3:8b chat model, about 5 GB in total.
  • An entry in the desktop's application menu and a shortcut on the desktop.

Nothing is exposed to the internet. Ollama, Qdrant and the app's own background service each listen on the server itself only (127.0.0.1), and the installer opens no firewall ports. Only you, from inside your desktop session, can reach them.

Opening the app

  1. Connect to your desktop over Remote Desktop, using the details on the desktop's Apps tab card.
  2. Double-click the Arynwood MCP shortcut on the desktop, or open the application menu and look under Development for the entry named arynwood-mcp.

The shortcut is placed for the desktop's login account. If your desktop shows it with a plain file icon or asks whether you trust it, choose to allow launching (right-click the icon, then Allow Launching); the application menu entry always works.

Troubleshooting

The app crashed and will not start again, or says its backend is already running: This is a known issue. If the app crashes instead of closing normally, its background service keeps port 8010 held, and the next launch cannot start. Open a terminal inside your desktop (or connect over SSH) and run:

pkill -f arynwood-backend

Then open the app again. Closing the app normally does not cause this.

Chat does not respond, or says a model is missing: If a model download could not finish during the install (for example after a network interruption), the install still completes and the card says so. Run these on the server:

ollama pull nomic-embed-text
ollama pull hermes3:8b

Replies are very slow: This is expected on a VPS, which has no GPU. Responses come from the CPU, and larger models are slower. Close other heavy programs in your desktop session, and consider upgrading to a plan with more vCPU.

Checking that the services are running: systemctl status ollama qdrant. To confirm they only listen locally, run ss -tln and check that ports 11434, 6333 and 6334 show 127.0.0.1.

The install was refused: The card gives the reason. The usual one is that no desktop environment is installed yet, or that the plan is below 4 vCPU or 8 GB RAM.

Removing Arynwood MCP

Click Remove on the Arynwood MCP card on the Apps tab. This removes the app, its desktop shortcut, and Qdrant together with its data, which includes your knowledge base index, so export anything you need first. Ollama is removed only if this install added it. If Ollama was already on your server, it is left in place and only the models this install downloaded are removed. Your desktop environment is not touched.

Limitations

  • KVM plans only; not available on LXC or on servers without a desktop environment.
  • Chat, knowledge base and design canvas only. Anything that needs an NVIDIA GPU is unavailable.
  • CPU inference is slow, and slower with larger models.
  • Arynwood MCP is alpha software from a third party. For problems inside the app, use its GitHub project. For problems with the server or the install, contact RackWorks support with your service ID.

¿Te ha resultado útil esta guía?

¡Gracias por tus comentarios!