This is great. Homelab AI feels like it's going to fun as heck. I currently have Claude maintain my homelab across all devices; it made homelab setup and maintenance go from "This is a trap that will fascinate you for years but never fully work right and waste time that would have better been spent elsewhere" to "This is actually a great idea and really extends my capabilities."
I've been doing something pretty similar, except instead of having a persistent opencode server, I've been using this workflow that runs opencode inside of the Forgejo action runners:
Still tinkering with it, but the gist is that I can invoke Opencode with /oc inside of an Forgejo issue, then it will come back with a PR for me to review.
Some times I feel like a lot of people in tech independently go through the same things right around the same time with few people writing/sharing about it.
I am also creating this and enjoyed the post and comments all going through the same thing :)
I think it’s because people in tech expect everything for free.
I had a conversation with my lawyer and I had “just one more question” that was going to take more than the time we had left in the current meeting. He said “schedule another 30 and let’s talk about that.”
I've been trying to find the motivation to do a write up on my AI lab, and this is just what I needed. Thanks for sharing. My setup is a similar idea, just with n8n/git/argo/k3s. It's mainly for automated workflows that Qwen or Gemma4 can handle.
Im doing something very similar. Running my OpenCode on a proxmox lxc. I have an additional layer of Kimaki, which gives you Discord integration (hate it or love it). Chatting with your codebase (voice messages, too, if that’s your jam), is very very cool.
> I set up OpenCode Web UI with Git access to make my homelab easier to manage. OpenCode pushes to Git, I approve the PRs, GitOps deploys the changes. Best of all, OpenCode runs as a server with persistent coding sessions synced across devices.
> I’ll share my homelab setup soon. There are about a dozen docker compose stacks for the services that I manage.
That is probably neat, but before I read, how many thousands of dollars would I need to spend to acquire the RAM and GPUs needed to do something similar?
Very cool, we're doing similar except we let agents open PRs as well + we track release metadata and agentic sessions via our ReARM system + we've recently launched an option for agents to track helm-based deployments via ReARM - https://docs.rearmhq.com/workflows/devops.html
I didn't mention this part, but while writing this I realized I could easily add a skill to hit the Forgejo PR API. There's no forgejo CLI like there is with GitHub sadly.
Do you use this at work or is it for vibe coding? Also, I don’t quite understand the problem you are solving. The solutions is a lot of technical parts put together, but why?
I see a lot of people using Komodo for it, though if I had to pick I'd go with Doco CD[0]. You can also use standard Ansible for just cron+bash script to git pull.
On the Podman side, I wrote a tool named Materia[1] for it, but there's also the wonderful Ansible quadlet role as well as Quadit and Orchess.
I recently setup Arcane and started migrating stuff from Truenas apps, they were all deployed as custom docker compose services so it worked out. Arcane supports Git syncs to auto deploy compose stacks, https://getarcane.app/docs/features/projects#sync-from-git
I'll write up some posts on my full setup soon.
so, the project is pretty much vibe coded, including the docs. It makes a lot more sense if you play around with it. It's just a docker host management UI, I like using it. It has gitops built in and a nice container log view. It doesn't do rollbacks, it only seems to sync from git and run compose up.
So first post in the blog, and it went directly HN frontpage.
Then, I said homelab AI, I thought it's an interesting post about local GPU setup (and I am really interested in this topic).. but no, just another hype post about how to use whatever-code...
I looked into running local models last month. They just aren't quite there for agentic tool use workflows without spending a small fortune. I'm hopeful smaller local models get much better soon.
I was also hoping to put out another post on my homelab setup, it has some neat stuff, but I haven't had a chance to finish it.
I think it heavily depends on what you're asking the model to do. Qwen3.6, both 27B and 35B-A3B, do agentic tool use very well. Their decision making is sus, but the dense model is decent in that way. A 4-bit quant for either of those can run on many home systems with a bit of configuration.
The biggest issue I've noticed is that the chat templates for open models are really hit or miss. The default Qwen3.6 chat template mostly works these days, but depending on your workload it may cause major issues. There are plenty of "fixed" chat templates on hugging face, but people report mixed success. It really seems to depend a lot on what the tool you're using expects.
My workflow is too different right now (gradually constrained to network less builds for reasons) but I am really enjoying how zeds agents have worked out in the past few weeks.
I have 27b, 35B-A3B and a cpu backed gpt-oss configured and use them in parallel, checking if one is getting ratholed and adding context or manual fixes.
I had various other systems setup and commercial models but really don’t use them.
It may be too interactive for some people, but it is a good mix of fail fast and often the places qwen3.6 was failing was eventually problems with the frontier models.
And this is with the unsloth defaults and hardened llama.cpp podman containers.
I do sometimes load other models or honestly just feed things into google’s free agent. But that is rare and to be honest manually fixing is typically faster and less error prone
This is great. Homelab AI feels like it's going to fun as heck. I currently have Claude maintain my homelab across all devices; it made homelab setup and maintenance go from "This is a trap that will fascinate you for years but never fully work right and waste time that would have better been spent elsewhere" to "This is actually a great idea and really extends my capabilities."
I've been doing something pretty similar, except instead of having a persistent opencode server, I've been using this workflow that runs opencode inside of the Forgejo action runners:
https://codeberg.org/dragonfyre13/forgejo-opencode
Still tinkering with it, but the gist is that I can invoke Opencode with /oc inside of an Forgejo issue, then it will come back with a PR for me to review.
Nice! I've done that with Claude Code + Forgejo but as a small separate app to run Claude in Docker: https://github.com/smithy-ai/smithy-ai
Some times I feel like a lot of people in tech independently go through the same things right around the same time with few people writing/sharing about it.
I am also creating this and enjoyed the post and comments all going through the same thing :)
I think it’s because people in tech expect everything for free.
I had a conversation with my lawyer and I had “just one more question” that was going to take more than the time we had left in the current meeting. He said “schedule another 30 and let’s talk about that.”
Fair!
I've been trying to find the motivation to do a write up on my AI lab, and this is just what I needed. Thanks for sharing. My setup is a similar idea, just with n8n/git/argo/k3s. It's mainly for automated workflows that Qwen or Gemma4 can handle.
Im doing something very similar. Running my OpenCode on a proxmox lxc. I have an additional layer of Kimaki, which gives you Discord integration (hate it or love it). Chatting with your codebase (voice messages, too, if that’s your jam), is very very cool.
That's very cool. Thanks, I'll have to check that out.
> I set up OpenCode Web UI with Git access to make my homelab easier to manage. OpenCode pushes to Git, I approve the PRs, GitOps deploys the changes. Best of all, OpenCode runs as a server with persistent coding sessions synced across devices.
> I’ll share my homelab setup soon. There are about a dozen docker compose stacks for the services that I manage.
That is probably neat, but before I read, how many thousands of dollars would I need to spend to acquire the RAM and GPUs needed to do something similar?
0? OpenCode is just a harness, it can connect to any model hosted online.
Very cool, we're doing similar except we let agents open PRs as well + we track release metadata and agentic sessions via our ReARM system + we've recently launched an option for agents to track helm-based deployments via ReARM - https://docs.rearmhq.com/workflows/devops.html
I didn't mention this part, but while writing this I realized I could easily add a skill to hit the Forgejo PR API. There's no forgejo CLI like there is with GitHub sadly.
There is, but it's limited. For example Forgejo does not expose CI build logs via the API so it's hard to make Claude auto-fix a build issue.
I still need to find the time to get into the Forgejo code and add that endpoint.
There is: https://forgejo.org/docs/latest/admin/command-line/
That's not an API tool. It performs direct database access for administrative functions on the Forgejo server.
But there is a different tool that is an API accessing CLI: https://codeberg.org/forgejo-contrib/forgejo-cli
That seems like a problem an LLM could solve. ;) (Assuming Forgejo has a reasonable REST/whatever API.)
How do you run inference for Open Code? What models are you running
Do you use this at work or is it for vibe coding? Also, I don’t quite understand the problem you are solving. The solutions is a lot of technical parts put together, but why?
I wonder how gitops is done with docker compose
I see a lot of people using Komodo for it, though if I had to pick I'd go with Doco CD[0]. You can also use standard Ansible for just cron+bash script to git pull.
On the Podman side, I wrote a tool named Materia[1] for it, but there's also the wonderful Ansible quadlet role as well as Quadit and Orchess.
[0] https://github.com/kimdre/doco-cd
[1] https://primamateria.systems or https://github.com/stryan/materia
I recently setup Arcane and started migrating stuff from Truenas apps, they were all deployed as custom docker compose services so it worked out. Arcane supports Git syncs to auto deploy compose stacks, https://getarcane.app/docs/features/projects#sync-from-git I'll write up some posts on my full setup soon.
What _IS_ Arcane? I fail to understand from their website https://getarcane.app/
Is it a deployment automation platform where it can run a project’s docker services, with rollback and all?
so, the project is pretty much vibe coded, including the docs. It makes a lot more sense if you play around with it. It's just a docker host management UI, I like using it. It has gitops built in and a nice container log view. It doesn't do rollbacks, it only seems to sync from git and run compose up.
So first post in the blog, and it went directly HN frontpage.
Then, I said homelab AI, I thought it's an interesting post about local GPU setup (and I am really interested in this topic).. but no, just another hype post about how to use whatever-code...
I looked into running local models last month. They just aren't quite there for agentic tool use workflows without spending a small fortune. I'm hopeful smaller local models get much better soon.
I was also hoping to put out another post on my homelab setup, it has some neat stuff, but I haven't had a chance to finish it.
I think it heavily depends on what you're asking the model to do. Qwen3.6, both 27B and 35B-A3B, do agentic tool use very well. Their decision making is sus, but the dense model is decent in that way. A 4-bit quant for either of those can run on many home systems with a bit of configuration.
The biggest issue I've noticed is that the chat templates for open models are really hit or miss. The default Qwen3.6 chat template mostly works these days, but depending on your workload it may cause major issues. There are plenty of "fixed" chat templates on hugging face, but people report mixed success. It really seems to depend a lot on what the tool you're using expects.
My workflow is too different right now (gradually constrained to network less builds for reasons) but I am really enjoying how zeds agents have worked out in the past few weeks.
I have 27b, 35B-A3B and a cpu backed gpt-oss configured and use them in parallel, checking if one is getting ratholed and adding context or manual fixes.
I had various other systems setup and commercial models but really don’t use them.
It may be too interactive for some people, but it is a good mix of fail fast and often the places qwen3.6 was failing was eventually problems with the frontier models.
And this is with the unsloth defaults and hardened llama.cpp podman containers.
I do sometimes load other models or honestly just feed things into google’s free agent. But that is rare and to be honest manually fixing is typically faster and less error prone
you can't explain the HN hug. You feel it, or your servers do.
Really cool! Do you autoapprove edits or do you approve manually?
I'll verify the PR code myself before merging, but that's usually a quick skim.