Self-Hosted Agent Factories and Plex Rollbacks

Homelab Highlights for 2026-08-21: self-hosting and homelab notes worth knowing.

[Audio embed placeholder]

News Roundup

Problems with hotio’s latest Plex release-727d8ac?

Users report that updating to the hotio Plex container release 727d8ac causes bandwidth warnings during local playback; rolling back to the previous release fixes the issue.

Why it matters: A reminder to pin container versions and check logs when updates cause unexpected behavior in media servers.

Source: GitHub

Self-Host Weekly (21 August 2026)

A weekly digest covering self-hosting news, including a large-scale hard drive reliability study, new tools like Spotify Soloist, and community updates like Debian’s vote on AI contributions.

Why it matters: A curated source for staying updated on the broader self-hosting ecosystem, tool releases, and infrastructure trends.

Source: selfh.st

Running AI agents in GitHub Actions with Docker Sandboxes

A guide on using Docker Sandboxes (microVMs) to securely isolate AI agents within GitHub Actions, allowing them to run tests and fix code with a private Docker daemon.

Why it matters: Applies the sandboxing concept from the lead story to CI/CD, offering a vendor-supported method for safe agentic automation in pipelines.

Source: Docker

Three ways to smuggle SQLite into Nix

A technical exploration of methods for embedding SQLite within the Nix package manager ecosystem.

Why it matters: Highlights the creative solutions often needed when integrating specific dependencies into declarative, reproducible build systems.

Source: Farid Zakaria

Decayfmt – a file format that corrupts itself a little every time you open it

An experimental file format designed to intentionally degrade its own data each time it is accessed.

Why it matters: A conceptual or artistic project that underscores the importance of data integrity and immutable storage in real-world systems.

Source: GitHub

Tool & Software Highlight: Building an (almost) fully self-hosted, sandboxed, agentic software factory

A detailed walkthrough of a system where AI agents work on code within isolated Docker containers and microVMs, running primarily on local hardware with local models via Ollama.

Why it matters: It provides a practical reference architecture for experimenting with agentic workflows while keeping compute and data on-premises, highlighting the crucial role of sandboxing for security.

Source: Jake Saunders

One Thing to Try

Sources

Transcript

Host A: Welcome to Homelab Highlights, the show that surfaces practical homelab wins and useful self-hosted tools.

Host A: Jake Saunders published a detailed walkthrough of what they call an almost fully self-hosted, sandboxed, agentic software factory. Host B: The concept is straightforward: AI agents work on code while everything runs on local hardware, with the key being the sandbox. Host A: The setup uses Docker containers and microVMs to create isolated environments where each agent gets root access inside but cannot touch the host. Host B: That isolation boundary is what makes it practical for homelabs; the blast radius stays inside the disposable environment. Host A: The blog describes connecting local models through Ollama, providing git tooling, and running a private Docker registry. Host B: It also mentions using Open WebUI as the interface, and the entire system runs on a single server. Host A: The post walks through networking, how agents persist their work, and where the “almost” in “almost fully self-hosted” comes from. Host B: There’s still a dependency on external AI providers for some tasks, but the core infrastructure stays on-premises. Host A: For operators wanting to experiment with agentic workflows while keeping compute local, this looks like a usable reference architecture. Host B: The tradeoff is that you’re responsible for managing those security boundaries yourself. Host A: Understanding exactly what the sandbox allows is crucial before letting an agent loose. Host B: [thoughtful] It’s a solid starting point, but the devil is in the details of the isolation policy. Host A: The post includes specific Docker commands and configuration examples, so it’s not just theoretical. Host B: That practical detail is what makes it useful for someone replicating the setup.

Host B: Quick updates from around the self-hosting world. Host B: First, the hotio Plex container release 727d8ac is reported to cause bandwidth warnings on local playback; rolling back to 664ca4c fixes it. Host A: Pin that older version for now and check server logs if you see the issue. Host B: Tom’s Hardware published a peer-reviewed study of 443,000 Backblaze drives over 1.66 million drive-years, ranking HGST most reliable and Toshiba least. Host A: Seagate and Toshiba failed at roughly twice the rate of WD and HGST, useful data for storage planning. Host B: Spotify released Soloist, a headless Connect client for Raspberry Pi and Linux, compatible with Free and Premium accounts. Host A: Could be a neat addition to a minimal audio setup, like a Pi Zero with a DAC hat. Host B: Debian opened a General Resolution vote on handling AI-assisted contributions, which could set policy for open source projects. Host A: How they rule may influence other distributions. Host B: Docker published a guide on running AI agents in GitHub Actions using Docker Sandboxes, isolating each agent in a microVM with its own Docker daemon. Host A: It’s the same isolation pattern from the lead story, applied to CI; the vendor’s tool is one option, but the concept is broader. Host B: And finally, someone built decayfmt, a file format that corrupts itself each time it’s opened. Host A: Either performance art or the worst backup strategy ever. Host B: [quick aside] It’s memorable, but probably not something you’d want to rely on.

Host A: Floppy is a self-hosted media tracker that appeared in Self-Host Weekly’s content spotlight. Host B: It’s positioned as an alternative to Trakt, Letterboxd, and TV Time, but runs on your own infrastructure. Host A: It unifies movies, TV, anime, books, games, music, and podcasts in a grid-based interface. Host B: Features include watch history, recap statistics, shareable lists, and owned-media tracking. Host A: The key for homelab setups are the sync integrations with Plex, Jellyfin, and Audiobookshelf. Host B: Those connectors let viewing history flow automatically from your media server to the tracker. Host A: Deployment is via Docker, requiring a separate Redis container. Host B: Adding it means one more container, but keeps your data private versus third-party services. Host A: The project is new, so expect some rough edges. Host B: Checking the source will show how well those integrations work. Host A: [curious] It fills a gap, but the quality of the connectors will define the daily experience.

Host B: A Reddit user gave Claude a domain and asked it to build whatever it wanted. Host A: Two weeks later, the agent had created 1f916.ai, a site where AI agents can register and interact, drawing 1.2 million views. Host B: The user finally checked the Cloudflare bill, and the costs were noticeable. Host A: This is one person’s experience, but it highlights the gap between a local experiment and a public service. Host B: Compute, bandwidth, and API costs can add up quickly when traffic spikes. Host A: For homelab operators moving projects online, it’s a reminder to monitor cloud dashboards. Host B: Auto-scaling and usage-based billing can lead to surprises. Host A: The experiment itself is fascinating, but the infrastructure aftermath resonates with anyone who’s managed servers. Host B: The takeaway is to instrument cost monitoring from the start, especially when mixing self-hosted components with external APIs. Host A: [skeptical] It’s a useful cautionary tale, but not a reason to avoid experimentation.

Host A: That’s Homelab Highlights for Friday. Until next time, happy hosting!