My Local AI Stack, Part 1: The Machine in the Room
Everything I make with AI runs on a computer in a room in my house. One machine, bought and paid for, doing work most people rent by the month. This is what it is and what runs on it.
Everything I make with AI runs on a computer in a room in my house. Not a data center. Not a subscription. One machine, bought and paid for, doing work that most people rent by the month. This is the first post in a short series about that machine and the software on it. Part 1 is the machine itself, and the next two posts cover the tools and then the pipeline.
The machine is not new and it is not fancy. That is the point.
- CPU: Intel i7-11700K.
- GPU: ASUS RTX 2080 Super with 8GB of video memory.
- RAM: 32GB of DDR4.
- Operating system: Windows 11.
- Drives: C holds the operating system, A holds the local AI work, P holds projects.
Eight gigabytes of video memory is not much by 2026 standards. It is enough. The whole stack is built around that limit. Small models run on the card. Big jobs get split up or run on the CPU. Nothing here needs a data center, and nothing here phones home to one.
The drives are split on purpose. C holds Windows and nothing else. A holds the local AI work, every model and every workflow. P holds projects, so a reinstall never takes the work with it. That split has saved me more than once.
Here is what runs on it, and what each piece does.
Dify is the glue. It is the workflow and agent layer, local software that builds workflows, connects to APIs, manages prompts, and chains agents together. Sixteen containers run when it is active. It ties everything else together.
llama-swap is the model router. It runs on port 8686 and swaps between local models based on demand. The big model running through it is Qwen3-35B-A3B, running on llama.cpp.
Ollama is the second model runner, on port 11434. It handles lighter models and tool integrations, the quick jobs that do not need the big model.
VoiceStudio is the voice tool. Voice cloning, dubbing, cinematic translation, and transcription, all on the GPU. It serves on port 3901 and runs version 0.5.9-beta.5. This is the tool behind the GridIowa narration and the audiobook work.
VoiceBox is the second voice tool. It also clones voices, and it ships a built in MCP server on port 17493, which means AI agents can be wired to speak responses out loud in a cloned voice. It has its own audio effects and a timeline editor.
ComfyUI is the image and video tool. It runs as a standalone desktop app and handles text to image work and the LTX video workflow.
Transcription runs through VoiceStudio, using Whisper. I pull transcripts from my own videos and from competitor videos to study how scripts are put together.
Recording is OBS for anything planned, and Nvidia Shadowplay instant replay for the moments nobody planned. Audio cleanup is Audacity. Video editing is DaVinci Resolve. Thumbnails are Canva, with a consistent template system and no AI generation. Code is Cline CLI with VS Code. Open WebUI, on port 3000, is the chat interface for the local models. Stirling PDF handles PDF chores. Notion tracks affiliate programs and project notes. Obsidian is the knowledge base.
The pipeline, in brief, is record in OBS, transcribe in VoiceStudio, let Dify flag the short clip moments, cut the clips in DaVinci Resolve, let Dify write the titles and descriptions, then publish. Part 3 will walk through that in full.
Not everything I tried stayed. Three tools got cut, and each one earned it.
- LocalAI is dropped. Version 4.7.1 had backend bugs, a rope theta error in Chatterbox, and a missing Fish Speech module. I will look again when it is patched.
- Seelen UI is avoided permanently. It corrupted the Windows PATH on install and broke WSL and Docker. Some tools do not get a second chance.
- n8n is shelved. Dify covers the same ground for this use case. I will revisit in three months if Dify shows gaps.
The software on this machine costs me nothing a month. That is the whole idea. A regular person with a used graphics card and a spare drive can run the same stack and keep the data home, with no subscription in the middle. Part 2 goes through the tools one by one, starting with what Dify actually does all day.