5 New Open-Source Repos to Self-Host This Week
Every Tuesday I go through my GitHub trending feed with a coffee and a fair amount of skepticism, because most "revolutionary new repo" posts are three people and a README. This week was different. I found five projects that are either already useful in a client's homelab or worth watching closely if you run your own infrastructure. If you're new to any of this, my self-hosting for beginners guide is a good place to start before you install anything below.
1. firecrawl/anydoc: turn your document graveyard into Markdown
Every homelab I've touched has the same problem: a folder full of old DOCX files, exported CSVs, PowerPoint decks from 2019 and PDFs nobody wants to open again. anydoc converts all of that (Word, PowerPoint, Excel, OpenDocument, RTF, EPUB, CSV, PDF) into clean Markdown. It's written in Rust with Node.js and Python bindings, MIT licensed, and it's part of the Firecrawl ecosystem, which already has a solid reputation for self-hostable scraping.
It launched on August 3, 2026 and has already crossed 18,100 stars, which makes it the fastest-growing repo I've seen this month by a wide margin. The docs are at firecrawl.github.io/anydoc. Clean Markdown means your documents become searchable, diffable in git, and ready to feed into any local LLM setup without the usual OCR headaches. I'm already planning to run this against a client's Nextcloud document library this week.
2. kimi-k3-in-c: a 2.78-trillion-parameter model on your CPU
kimi-k3-in-c is one of those repos that sounds fake until you read the code. It's a from-scratch inference engine, written in portable C99, that runs the full 2.78-trillion-parameter Kimi K3 model on a single CPU using just 8.24 GB of RAM. No BLAS, no framework, no GPU required. It uses linear attention, which is the trick behind the tiny memory footprint. Apache-2.0 licensed, created August 1, 2026, already past 6,300 stars.
Let's be honest about what this is and isn't. It will not replace llama.cpp for anything you're serving to real users; you're talking tokens per minute here, not tokens per second. But it proves frontier-scale weights are becoming portable to hardware that ordinary people own, which is exactly the theme I wrote about in last week's Muse Glimmer post. There's a companion project worth knowing about too: sqliteai/warp, Apache-2.0, written in C, over 2,200 stars, which runs the same full model by streaming activated weights straight off NVMe instead of loading them into RAM. Same race, different bet on where your bottleneck lives. This is exactly the "run it on a 200 dollar mini PC" energy that self-hosting is supposed to be about.
3. CopilotKit/OpenBot: agents with their own computer
OpenBot gives each AI agent its own browser, its own files and its own tools, with every action decided before it happens and logged after. It supports AG-UI agents and MCP, is written in TypeScript, MIT licensed, and just went public on August 17, 2026, so this is genuinely fresh: it already has over 2,500 stars in its first week. There's more detail on the CopilotKit site.
Agents that can actually browse and click things are usually locked inside a SaaS product with a monthly invoice attached. OpenBot is an attempt to make that self-hostable, and the "decide first, record after" design is exactly the right instinct if you're the one responsible for what an agent does on your own hardware. It's a young project and I wouldn't hand it production access to anything sensitive yet, but if you're comfortable spinning up containers, my Docker for beginners guide covers the mistakes worth avoiding before you let any agent run loose in one.
4. genspark-ai/genoffice: a free AI office suite
genoffice is a free, open-source AI office suite for macOS, Windows and Linux. It handles Word, Excel, PowerPoint, PDF and Markdown editing with built-in AI agents baked into the app. It's an Electron app written in TypeScript, Apache-2.0 licensed, created July 31, 2026, and already sitting past 3,500 stars. More at genoffice.ai.
Worth a clarification here: this is a desktop app, not a server, so "self-host" in this case really means "run your own copy and stop paying a subscription" rather than anything you're putting in Docker. That still counts in my book. It's the most interesting open-source answer I've seen so far to the "office suite plus AI" question, and for freelancers and small teams tired of per-seat pricing, it's worth a real trial.
5. Obiente/nc-native: a native Nextcloud client, finally
I install Nextcloud for clients almost every month. It's one of my steadiest Fiverr gigs, and the client ecosystem has needed a modern native app for years. nc-native is an adaptive native client for mobile and desktop, written in Kotlin, AGPL-3.0 licensed, with dynamic support for the wider Nextcloud app ecosystem. It was created July 24, 2026 and is still getting active commits as of today.
At 478 stars it's small, and it's an early-stage project from a small team, so this isn't a "replace your sync setup tonight" recommendation. But the idea is sound, and if you've read my Nextcloud vs Google Drive breakdown, you know how much the client experience matters when you're trying to convince a client or a family member to leave Google Drive behind. I'm watching this one closely and I'll report back once I've actually deployed it somewhere.
Which ones made my shortlist?
If I had to pick two to actually deploy this week, it's anydoc for cleaning up document chaos and genoffice for anyone who wants off subscription office software. kimi-k3-in-c I'm keeping around purely as a proof of concept, not a daily driver. OpenBot and nc-native are both promising but early, so they go on the watch list rather than the production list.
That's my read for this week. If you've tried any of these already, or found something I missed, drop a comment, I'm always looking for the next thing to break on a test server before I trust it with a client's data.