스크립트
00:00:00The most starred new repo of the summer, over 83,000 stars in six weeks.
00:00:05It didn't come from Google, Microsoft, or OpenAI. It came from PewDiePie. Yeah, that PewDiePie.
00:00:12And if you open the issue tracker, the most upvoted issue says, I quote,
00:00:17this is pure har. And another comment saying, I'm literally scared to even run this on my machine.
00:00:22So I'm going to do it for you. And we're going to see what this even is.
00:00:30The project is Odysseus and calling it another chat UI sort of undersells it. It's a self-hosted
00:00:38AI workspace that does apparently all of these things that are way too long for me to even read.
00:00:44While doing all this, it apparently survives between sessions, one app running on your machine.
00:00:50A GPL license, no cloud tier, no subscription. Every chat and email sits in a plain local data
00:00:56folder you can backup yourself. And the backstory explains everything. Last year, PewDiePie fell
00:01:02down the local AI rabbit hole probably harder than most of us. He built a 10 GPU rig with a modded 48
00:01:09gigabyte 4090s 70 billion parameter models and a homemade UI called chat OS with a council of bots
00:01:18that voted on answers until they started colluding against him. Odysseus is that system rebuilt for the
00:01:24rest of us. He shipped on May 31st, 30,000 stars in 48 hours. We're going to see what this really is.
00:01:30If you enjoy coding tools to speed up your workflow, be sure to subscribe. We have videos coming out all
00:01:35the time. So let's run this because on a Mac, it's genuinely one script clone run Docker and it sets
00:01:41everything up. When I grep, I'll get back a temporary password in the terminal and it's served on port 7000.
00:01:48I can log in with admin and that password. And now we're looking at this mono spaced red on dark
00:01:52terminal styled interface. XDA called it weirdly great while a hacker news commenter called it
00:01:58atrocious. And honestly, that split tells us everything about this project.
00:02:03First stop is the cookbook. It scans this Mac, the chip, the unified memory, and tells me which local
00:02:09models will actually run here, then downloads and serves them through Llama CPP with a full metal
00:02:16acceleration. That one screen replaces an hour of digging around. I can drag the boxes around,
00:02:22which is cool. But the next interesting thing here is comparing models. One prompt,
00:02:27three local models, side by side, blind testing. I can then enter my prompt. And when it's done,
00:02:34I can pick the winner before it reveals which model wrote that. This is the bot council. That's kind
00:02:40of cool. It's all built on a fast API app for containers, Odysseus, chroma DB for vector memory,
00:02:47a bundled Seer XNG search engine, and Nifty for push notifications. That bundle is why Deep Research,
00:02:54which I'll talk about in a minute, works with zero paid APIs. It searches through its own Seer XNG,
00:03:00reads the sources, and streams back a cited report all local. The memory is real. Tell it a fact today,
00:03:06asks tomorrow in a fresh session, it knows. The email client triages your actual inbox over IMAP
00:03:11and drafts replies. I didn't want to sync my email to this for obvious reasons, but it's built in.
00:03:17There is an integrated calendar where you can add things just like a normal calendar, but now the LLM
00:03:22can actually read it. I can actually go into deep research here, which I'm circling back to.
00:03:26I can ask it a question and it runs a multi-step agent loop. The best analogy I've got, self-hosted
00:03:33cloud projects with an inbox stapled onto it. Now, one thing I want to be fair about, the agent loop,
00:03:39cookbook, the research pipeline are adapted with proper credit from OpenCode, LLM Fit,
00:03:46and Alibaba's Deep Research. This is just skilled assembly. It's not an invention from scratch,
00:03:52and honestly, it doesn't need to be. Which sets up the question Hacker News asked word for word.
00:03:58Why not just use OpenWebUI? Totally fair. OpenWebUI and LibreChat have years of maturity in actual
00:04:05version releases. Odysseus wins on vastness. Nobody else ships email calendars, scheduled agents,
00:04:11push notifications, and a hardware-aware model manager in one box. There's a licensing wrinkle,
00:04:17too. OpenWebUI's license won't let you remove its branding, which Hacker News pointed out isn't fully
00:04:23open source. Odysseus is plain AGPL. But if you want chat plus RAG and that's all you need,
00:04:30just take another tool. There's no argument there. Now, the part you actually need to hear,
00:04:34because I'm not pushing this project by any means, that top issue says WTF is going on here.
00:04:41It describes 800 merged pull requests of LLM Slop, a 30,000-line CSS file, random inline JavaScript,
00:04:49and right now, close to 900 PRs sit open, many written by AI, others submitted by fans.
00:04:56There are zero tagged releases. Zero. You're running a moving dev branch, un-penned dependencies,
00:05:03install it today, and tomorrow you get two different apps. The stars are partly fandom,
00:05:07obviously. The man has 110 million subscribers, but stars are not code review. And read the project's
00:05:14own threat model. The agent gets shell access, file access, and the ability to send email. Their words,
00:05:21treat it like an admin console. Never expose this to the internet. A Mac-specific catch in Docker,
00:05:27inference only runs CPU only, because Docker and Mac OS can't touch the metal GPU. So,
00:05:32my verdict is two verdicts. As a local playground, it's cool. As a production dependency, absolutely not.
00:05:39And it's fine to be both. But here's what the it's-just-slop crowd is missing. Over 300 people
00:05:45have contributed to this now. The community forced through a code owner's file wrote a full architecture
00:05:52reform proposal that drew 100-plus comments. And two days ago, the repo moved into its own GitHub org.
00:05:59This thing is messy, crowded, uneven, and more alive than almost any project I've seen in a while. So,
00:06:05here's my takeaway. It's not about the code. The code is honestly the least important thing in that
00:06:09repo. What matters is that 110 million people just watch someone they say they trust say you can run
00:06:15your own AI on your own machine and own your own data. Self-hosting AI is going mainstream,
00:06:20but that doesn't always mean we should trust it. If you enjoy coding tips and tricks like this,
00:06:24be sure to subscribe to the BetterStack channel. We'll see you in another video.