Personal AI · Linux and Windows desktop alpha

Arynwood MCP

A personal AI that runs on your own machine.

Try it in your browser Download v0.5.0 View source
  • v0.5.0 alpha
  • Linux x86_64
  • Windows x64
  • AppImage · .deb · .exe
  • Source-available
Arynwood MCP chat window: the Arynwood persona explains that it can search the web and a knowledge base and drive a running Kdenlive editor, with persona tabs above the conversation
The real interface running in your browser with scripted data. No install needed.

Released October 10, 2026

What changed in 0.5.0?

The latest release makes Community work with a Grove on a server and every page readable, with security updates. Community inside the app, update notices, CPU mode, chat and image endpoints, and the October 9 MCP hardening are included.

  • Community with a Grove on a server: a Grove behind Cloudflare or nginx, such as community.arynwood.com, sends compressed answers, which the 0.4.9 Community page couldn’t read. Arynwood now unpacks them first.
  • Readable everywhere: a contrast check of every page found text below the 4.5:1 standard, including your own messages in Chat. All of it now passes.
  • Security updates: Python, frontend and Rust dependencies with published advisories are updated. npm audit, pip-audit and cargo audit report no known vulnerabilities.
  • Community inside the app (since 0.4.9): the Community page shows your Grove’s planner and calendar, tasks, lists, notes, board, discussion, members and invitations, with screens from Arynwood Grove 0.4.1. A Grove’s host also gets its IRC rooms, peer messages and Mail.
  • Choose your Grove: the one on your computer, or one on a server by its https:// address. Arynwood keeps the Grove sign-in and contacts only the Grove you chose, which receives what you post to it.
  • Update notices: check automatically at startup and every six hours while open. A dismissible banner opens the official release page in your browser; installation stays manual. Anyone on 0.4.8 or 0.4.9 sees the notice for this release; earlier versions need one manual install. Checks send no credentials or workspace content.
  • CPU mode: Auto, On and Off controls mark and disable NVIDIA-only tools. CPU recording and editing remain available. Local CPU chat can be slow; a remote endpoint is an option.
  • Chat and image endpoints: choose an Ollama or OpenAI-compatible chat server and an OpenAI-compatible image API for Design Center. Choices persist. Remote services receive the prompts and data you send.
  • MCP hardening: tool definitions are rechecked before dispatch; changed definitions, duplicate names and invalid arguments fail closed. External schema fetches are disabled, and common credentials are redacted from context and evidence.

Still a single-owner desktop alpha, without an OS sandbox or enforced offline mode. Windows is unsigned; clean-machine installation and optional Windows integrations remain unverified. Read the full release notes and updated security brief.

Download 0.5.0

Linux AppImage · Debian/Ubuntu .deb · Windows x64 installer

Verify downloads with Linux checksums or Windows checksums. Models and supporting services are installed separately.

What’s inside

One workspace, your models

Start with Ollama chat, then add the services your workflow needs. Models and supporting services are installed separately.

  • Chat and models

    Multi-persona chat that streams from Ollama. Browse, pull and delete models, and register local or remote Ollama and OpenAI-compatible servers.

  • Knowledge search

    Teach it from a URL, a file or pasted text, then search your own documents. Uses Qdrant and an Ollama embedding model.

  • Design canvas

    A canvas design tool. Image generation connects to a separately installed A1111 service.

  • Music and audio

    Optional music and audio tools through separately installed MusicStudio sidecars.

  • MCP tool integrations

    Model Context Protocol tool servers, including Kdenlive editing through Cutroom’s MCP server. Each tool server is configured separately.

  • Built for daily use

    A Ctrl+K command palette and a system status drawer on every screen, with live service health on the dashboard.

Experimental

A remote control for your own workspace

Off by default and run from a source checkout; the desktop packages don’t include it. It answers only you.

Gateway guide

  • Runs without the desktop window

    A background service you start yourself keeps your conversation going, and picks up where it left off after a restart.

  • Works through multi-step tasks

    Your agent can use a tool, read the result and choose the next step until the job is done. Anything that deletes or publishes waits for your approval, and is refused when nobody is there to approve it.

  • A memory you can read and edit

    Facts you tell it are written to a plain MEMORY.md file and dated notes on your disk, and it remembers them after a restart. Only your own messages read or write it.

  • Message it over IRC

    Talk to your agent from any IRC app, through your own IRC server. It answers only your services account, not your nickname, and ignores everyone else.

Where things stand

  • Shipped: the v0.5.0 desktop alpha for Windows and Linux, with Community inside the app, automatic update notices, CPU mode, chat and image endpoints, and MCP hardening
  • Alpha: Arynwood Grove, the separate service that keeps Community spaces
  • Experimental, off by default: the headless gateway

What leaves your machine

Arynwood checks GitHub for a newer stable release at startup and every six hours while open. The check sends no credentials or workspace content; GitHub receives normal connection metadata, including your IP address. There is no telemetry. Other data goes out through features you use: web search, learning from a URL, downloads, social publishing, deploys, IRC, a Grove on a server when you use Community, or a remote model server you add. The security notes list what is protected and what isn’t yet.

Requirements

What you need to run it

From the project documentation. Read the platform requirements and the Windows setup notes.

Operating systemLinux x86_64: AppImage or Debian/Ubuntu .deb. Windows x64 (alpha): a per-user installer; WebView2 is required and the installer downloads it if needed. macOS is not supported.
Memory16 GB RAM minimum in the project documentation. Model size affects memory use.
StorageAt least 20 GB free, plus space for models and generated media.
Chat GPUOptional. Ollama can run chat on a CPU, more slowly. See measured CPU speeds.
Creative GPU toolsAn NVIDIA GPU for the documented GPU services. At least 12 GB VRAM is the project’s practical recommendation. On Windows, these tools need their own Windows setup and haven’t been validated yet.

Get started

Start with one chat workflow

Get chat working first, then add knowledge search and creative tools one at a time.

Full setup guide

  1. Step 01

    Download and install

    Get the package for your system from the v0.5.0 release. On Linux, check the AppImage or .deb against SHA256SUMS.txt, then make the AppImage executable or install the .deb with its GStreamer packages. On Windows, check the setup .exe against SHA256SUMS-windows.txt and run it. The installer is unsigned, so Windows may warn you first.

  2. Step 02

    Add Ollama and a model

    Ollama isn’t bundled. Install it, then pull a model that fits your RAM or GPU memory. A 3B model is comfortable on a CPU alone.

  3. Step 03

    Check chat, then expand

    Pick the model and send a message. Once that works, add Qdrant and an embedding model for knowledge search, then any creative services.

# Linux: install Ollama
curl -fsSL https://ollama.com/install.sh | sh
# Windows: run the installer from https://ollama.com/download/windows

# Then, in a terminal or PowerShell, pull a starter model
ollama pull llama3.2:3b

Needs extra setup

  • Image generation: A1111 runs as a separate service
  • Music and audio: install the MusicStudio sidecars and their models
  • Kdenlive editing: the compatible editor, a running MCP server and local configuration
  • Social publishing: network access and platform accounts

Not in the packaged alpha

  • Chatrooms or end-to-end encrypted messaging
  • LoRA training, script-based GPU tools and the project file browser (source checkout only)
  • macOS builds
  • On Windows: the Linux-only tool installers and the local Community launcher. GPU tools, Kdenlive and MusicStudio are not yet validated there

Questions

Arynwood MCP questions

What it includes, what it needs, and what stays on your machine.

What is Arynwood MCP?

Arynwood MCP is a local-first AI workspace and personal AI agent for Linux and Windows. It runs language models on your own machine through Ollama and adds document search, memory, a design canvas and MCP tool integrations. The current release is the v0.5.0 desktop alpha.

What does Arynwood MCP send off my machine?

Arynwood checks GitHub for a newer stable release at startup and every six hours while open. The check sends no credentials or workspace content; GitHub receives normal connection metadata, including your IP address. There is no telemetry. Your chats, documents and memory stay on your disk, and the models run on your computer through Ollama. Other data leaves through features you use: web search, learning from a URL, downloading models and tools, social publishing, deploying to your server, IRC, a Grove on a server when you use Community, or a remote model server you add. There is no enforced offline mode, so the security notes list what is protected and what isn’t yet.

Can an AI assistant run 100% offline?

Local chat, document search and memory can work without an internet connection once the models are downloaded. Arynwood has no enforced offline mode. Automatic update checks, downloads, web search, remote model servers and social publishing use the network when available. Failed update checks stay quiet.

Can I talk to my home AI server from my phone?

Only experimentally. The headless gateway, which is off by default and runs from a source checkout, lets you message your own home agent from any IRC app on your phone, through your own IRC server. It answers only your services account and ignores everyone else. The model and your memory stay on your home machine; only the chat messages travel, encrypted with TLS.

Does it include the AI models?

No. The download contains the interface and backend. Ollama and its models are installed separately, and so are Qdrant, A1111 and MusicStudio if you use those features.

Does it run on Windows?

Yes, as an alpha. Version 0.5.0 includes a Windows x64 installer with the backend bundled. It is unsigned, and clean-machine installs, microphone and camera capture, downloads and the optional integrations still need testing. As on Linux, Ollama and Qdrant are installed separately. App data is kept in %LOCALAPPDATA%\arynwood-mcp and survives upgrades. Read the Windows setup notes.

Can I try it without installing?

Yes. The browser demo runs the real interface with scripted conversations. It is not a live AI; features that need a GPU or backend say so.

Do I need a GPU?

Chat can run on a CPU through Ollama. On a 2012 8-core Xeon with no GPU, a 3B model generated about 13 tokens per second, faster than most people read. See the full CPU benchmark. GPU generation features have separate NVIDIA requirements.

Which model does Arynwood MCP use for tool calls?

Hermes 3 8B by default, chosen by measurement. On a 12 GB RTX 3060 it made 27 of Arynwood MCP’s 30 routing decisions in 0.12 seconds each and followed no instruction planted in a tool result. See how seven models compared.

Is it open source?

Its source code is public on GitHub under the Arynwood MCP Source-Available License, which is not an open-source license. Read the LICENSE file for what it allows.

What are Arynwood Community Groves?

Invite-only, private spaces for families, friends, teams and small communities, with no ads and no sales. A Grove is the separate Arynwood Grove service, with no AI features, that keeps the spaces on your own computer or on a server your members reach. Since version 0.4.9, the Community page in Arynwood MCP shows your Grove’s calendar, tasks, lists, notes and discussion; use 0.5.0 or later with a Grove on a server. It is an alpha: the host’s peer-to-peer messages are not yet verified end to end.

Keep reading

Guides and related tools

Local AI setup & integration

Want it set up on your machine?

Tell me your Linux distribution, hardware and the task you want AI to help with. I’ll scope a paid setup around your machine.

Email for a quote See setup services

Running AI on your own Linux machine? Join the Linux Local AI Community on Facebook to compare setups and ask questions.