Everything you can run on a DGX Spark

After getting vLLM and VS Code Copilot running in part 2, I spent a few days going through everything else the DGX can run. The local AI ecosystem is enormous right now. Tools launching every week, repositories with five-digit star counts, every framework claiming to be production-ready.

I made a list and researched each one. This post is that list, with verdicts. The categories are inference, routing, coding agents, web UIs, vector databases, agent frameworks, image generation, voice, and personal tools. Most of it I skipped. Here is everything I looked at and what ended up on the machine.

Inference engines

vLLM is what you should default to on the DGX. It was built for NVIDIA hardware from the ground up. Everything else is a tradeoff: Ollama is simpler but slower, llama.cpp is portable but CPU-first, NIM/TensorRT-LLM is fastest but locked to the NVIDIA stack.

Routing

LiteLLM sits between every tool and the models. Every agent, every UI, every script points at LiteLLM and never touches vLLM directly. Swapping models or adding a tier is a config change. OpenRouter is useful on the Mac side for reaching paid cloud models alongside the DGX.

Coding agents

VS Code Copilot covers interactive in-editor work. It does not cover autonomous work: read this codebase, find the bug, open a PR. That is a different class of tool. OpenCode handles interactive terminal sessions, OpenHands handles async autonomous tasks. They are complementary.

Web UIs

Open WebUI is the default and I already had it running. AnythingLLM is worth adding if you want a turnkey "drop docs, query them" interface. The others are polished but redundant if you have one UI you like.

Vector databases

I already had LanceDB from OpenClaw's memory plugin and pgvector through Postgres. Neither hit a wall, so I added nothing else. If either does, Qdrant is the pick for serious self-hosted RAG.

Agent frameworks and orchestration

I looked at everything here. Most frameworks try to be the whole stack. The one I would actually install is n8n: workflow automation with AI nodes, 400+ integrations, the "GitHub webhook fires, DGX processes it, Slack gets a message" layer. The heavy frameworks (LangChain, Dify) are better suited to teams building products on top of LLMs than individuals running a personal stack.

Image generation

ComfyUI is the standard. Node graph, supports FLUX, WAN 2.2, and LTX-2, which are the validated models for the Spark. The others are fine but ComfyUI has won this category.

Voice and audio

I run faster-whisper on the DGX as an always-on transcription endpoint. Everything else in this category requires a specific voice use case I do not have yet.

Personal tools

This was the category that surprised me most. None of these require a DGX: any always-on home server would do. But since the machine is already running, they are free to add. These are also the use cases that make local AI actually worth it: things you would never hand to a cloud provider.

Immich is self-hosted Google Photos: face recognition, semantic search, everything you would pay Google for, running locally on your own hardware. Paperless-ngx handles documents with OCR: drop a receipt, contract, or tax scan and it finds it. Karakeep is self-hosted Pocket with AI auto-tagging. I use all three.

What I deliberately did not install

The trap with a machine this capable is that everything seems worth trying. The filter I kept using: does this solve a problem I have today, or one I am imagining? Most tools that seemed fascinating failed that test.

If you cannot name the specific workflow a tool enables, skip it for now. Install one thing, use it for a week, then decide. What is actually running on the DGX now is what I built SparkMonitor to track, and I will cover that in a future post.