@pinguy3ML Researcher, independent
Shipped: Rhizome Stack
A local-first AI agent stack that turns OpenClaw + Open WebUI into one coherent personal AI environment.
Rhizome combines local or hosted LLMs with persistent memory, agent skills, semantic retrieval, voice, a sandboxed code interpreter, and local image, music and video generation. It manages the glue between them — including safely handing limited GPU memory between chat and creative models, then returning the result directly to the conversation.
Built for Linux and WSL2, with an emphasis on owner-operated hardware, privacy, recoverability and avoiding unnecessary cloud dependencies. The current alpha.15 candidate exports a verified 227-file release, with automated regression, manifest and privacy checks.
Local AIAI AgentsMultimodalOpen SourcePrivacyLinux
Comments7
How does Rhizome recover GPU ownership if a creative task crashes or times out during the handoff from chat? A reproducible example showing chat → image generation → failure → chat recovery, including peak VRAM, would be useful for evaluating it on constrained hardware.
saved Open WebUI chat -> packaged Qwen, MiniMax or WAN tool -> shared generation lock -> refuse if ComfyUI or another media job is busy -> unload every Ollama runner with keep_alive=0 -> wait until Ollama /api/ps is empty -> start the selected packaged generator on demand -> submit generation exactly once -> retry only read-only status polling -> free ComfyUI models and stop the engine owned by that application -> wait until the ComfyUI port is closed -> copy the result into Open WebUI's upload store -> create the Open WebUI file row and attach it to the saved message -> return to model continuation; Open WebUI reloads the local chat model
What is the minimum, setup Time and cost for this, and what if something goes wrong?
cost electric bill. Time how ever fast the internet is to download the models. Hardware 32GB of RAM RTX 3060 12GB
How do you Safeguard Memory and How do you keep it persistent?
Is Rhizome going to stay free, or do you plan to charge for a supported version once it's out of alpha?
Love the owner-operated, local-first angle. I built something much smaller in the same spirit (Forge AI: answers from your Obsidian notes with citations, runs on Ollama). How do you decide when to swap the chat model out for an image/video model? That handoff seems like the hardest part on limited VRAM.
7 comments
How does Rhizome recover GPU ownership if a creative task crashes or times out during the handoff from chat? A reproducible example showing chat → image generation → failure → chat recovery, including peak VRAM, would be useful for evaluating it on constrained hardware.
saved Open WebUI chat -> packaged Qwen, MiniMax or WAN tool -> shared generation lock -> refuse if ComfyUI or another media job is busy -> unload every Ollama runner with keep_alive=0 -> wait until Ollama /api/ps is empty -> start the selected packaged generator on demand -> submit generation exactly once -> retry only read-only status polling -> free ComfyUI models and stop the engine owned by that application -> wait until the ComfyUI port is closed -> copy the result into Open WebUI's upload store -> create the Open WebUI file row and attach it to the saved message -> return to model continuation; Open WebUI reloads the local chat model
What is the minimum, setup Time and cost for this, and what if something goes wrong?
cost electric bill. Time how ever fast the internet is to download the models. Hardware 32GB of RAM RTX 3060 12GB
How do you Safeguard Memory and How do you keep it persistent?
Is Rhizome going to stay free, or do you plan to charge for a supported version once it's out of alpha?
Love the owner-operated, local-first angle. I built something much smaller in the same spirit (Forge AI: answers from your Obsidian notes with citations, runs on Ollama). How do you decide when to swap the chat model out for an image/video model? That handoff seems like the hardest part on limited VRAM.