Transcription
A few weeks ago, I showed you 10 GitHub repos so good they shouldn't be free and the exact paid products they replace. That video took off and the comments did something better. They kept naming bills I'd missed. So, this is part two. 10 more repositories none repeated from part one. Every number fact-checked this week. And this round, the names got bigger. PewDiePie shipped one of these. ByteDance shipped one. Alibaba open-sourced theirs days ago. Same rules as last time. For every repo, I'll show you the exact paid product it replaces and the honest catch nobody mentions. Let's go.
Number one, Odysseus. On May 31st, PewDiePie, yes, that PewDiePie, released his own self-hosted AI workspace and it pulled 20,000 GitHub stars in the first 24 hours. It's the ChatGPT experience running on your machine. Chat with local models through vLLM, llama.cpp, or Ollama. Or point it at OpenAI or OpenRouter with your own keys. And it goes past chat. Agents that take a task and run it with web, file, and shell access. Deep research that reads sources and builds a visual report. It even triages your email and drafts replies in your style. ChatGPT Plus is $20 a month. Pro is 200. This is that interface without the subscription. The honest catch, it's days old and moving fast. The default branch is literally called dev. But 66,000 stars later, it's the most watched repo of June.
Number two, Dear Flow. This one is from ByteDance, the TikTok company, and it's MIT licensed. It started as a deep research framework. Version two turned it into a full agent harness. Every task gets its own sandboxed computer. It spawns sub-agents on the fly, each with scoped context and its own tools. Skills load like slash commands. And here's the feature that got me. The message gateway. You can kick off a research run from Telegram or Slack, and just walk away. OpenAI's deep research at full quota lives on ChatGPT Pro, $200 a month, and plus users get a fraction of the runs. Dear Flow gives you the engine with no quota at all. The honest part, you bring your own model keys. Long runs burn real tokens, and the recommended self-host box is beefier than a laptop.
Number three, Voicebox. Jamie Pine, the developer behind SpaceDrive, built a local AI voice studio that clones a voice from a few seconds of audio. It runs seven different text-to-speech engines, including When three and Chatterbox. It speaks 23 languages. It ships a multi-track timeline editor for podcasts and conversations. There's a global dictation hotkey and MCP support, so your agents can literally speak in your cloned voice. ElevenLabs charges $22 a month for professional voice cloning, 99 on Pro, and meters you by the character. Voicebox is MIT licensed and meters nothing. The honest part, it's pre-version one. The last release was late April. And quality depends on which engine your hardware can run. The lightest one fits in about 1 GB of VRAM.
Number four, Open Notebook. This is Notebook LM self-hosted. Feed it PDFs, videos, audio, web pages, then chat with your sources with citations. It supports 18 plus AI providers, including Ollama running fully local. And the podcast feature actually beats Google's. Notebook LM is locked to two hosts. Open Notebook generates audio with one to four speakers, each with a custom profile. Now, the honest reframe on the price. Notebook LM's core features are free. What $20 a month of Google AI Pro buys you is higher limits. Open Notebook has no limits to raise. It runs in Docker in about 2 minutes. And your sources never leave your machine. The catch? It's one maintainer's project, and cloud models still cost API tokens. Free only truly means free when you run local models.
Number five, Meet Lily. An AI meeting note taker with no bot joining your call. It captures your mic and system audio locally. Transcribes live with Whisper or the four times faster Parakeet, and summarizes with Ollama. So, a meeting goes from audio to notes without ever touching the cloud. 234,000 downloads. MIT licensed. macOS and Windows. Otter charges about $17 a month, and caps your minutes. Fireflies is 18 per user. And both of them ship a bot into your meeting and your transcript to their servers. The honest part, speaker identification, who said what, has not shipped yet. It's planned for this month and it lands in the paid pro tier. But the free community edition is the full tool, not a trial.
Number six. Immich, the big one. 103,000 GitHub stars. The most starred self-hosted photo project on the planet. It's Google Photos on your own server. The phone app auto backs up your camera roll. Machine learning recognizes faces and objects. And you can search your library in plain English, red car at the beach. Because it runs clip locally. Shared albums, memories, even 360 photos. Google charges $10 a month for the 2 TB plan. A $100 a year, forever, to look at your own memories. Immich is free. The team works on it full time and version two is officially stable. The honest part, you supply the server and the drives. And even Immich tells you to keep a separate backup. Self-hosting your photos means you own the failure mode, too.
Number seven. Diad. The lovable experience. Describe an app, watch it get built. Except it runs entirely on your machine. Any model, your own keys, no platform lock-in, Mac and Windows. Lovable Pro is $25 a month, metered in credits. V0's team plan is 30. Diad's core is Apache licensed and free. You pay your model provider and nobody else. Version 1.3 shipped two days ago. So, this one is moving fast. The honest part, and you'll see this pattern across the list, it's open core. There's a $20 pro tier with extra credits and large code base modes. The free app is real and complete. The company makes its money on the add-ons. That's the deal, and it's a fair one.
Number eight, Multika. This is a manager for AI coding agents. You assign a GitHub issue to an agent the way you'd assign it to a colleague. It shows up on the board, posts comments, and reports blockers before you ask. You can group agents into squads with a leader that decides who picks up the work. It runs 11 different agent CLIs, Cloud Code, CodeX, Co-pilot, Gemini, Cursor. Their pitch, two engineers and a fleet of agents can move like 20. Compare that to Devin, the managed AI team made that launched at $500 a month. Today, its team plan is $80 a month plus $40 per seat before you pay for the agents themselves. Multika's platform fee is zero. You self-host it with Docker, and pay only for the agents you already use. The honest part, it's 5 months old, and the license is source available. Free to run for your own team, not to resell as a service.
Number nine, and this one is days old. On June 10th, Alibaba open-sourced Open Code Review, the AI code review assistant they ran internally. By their own count, it served tens of thousands of their developers, and flagged millions of defects over 2 years. The design is the interesting part. Deterministic pipelines handle file selection and rule matching. An LLM agent handles the judgment. Because, in their words, pure agents suffer incomplete coverage and unstable quality. It posts precise line-level comments. It ships fine-tuned rules for null pointers, thread safety, cross-site scripting, and SQL injection. And it speaks both OpenAI and Anthropic APIs. CodeRabbit charges $24 per developer per month for this exact job. This is Apache licensed. Bring your own model key and the per-seat bill disappears.
Number 10. Dokploy. Its GitHub description is literally open-source alternative to Vercel, Netlify, and Heroku. Self-host it on a $5 VPS and you get the whole dashboard. Deploy apps in any language. Native Docker Compose. Six databases out of the box. Postgres, MySQL, Redis, and more. Multi-server scaling with Docker Swarm. Real-time monitoring. Vercel Pro is $20 per user per month and the overage bills are their own genre of horror story. Heroku dynos run from $7 to $1,500 a month. Even Dokploy's own managed cloud is 450 per server, which tells you the margin everyone else is charging. The honest part, it's still version 0.29. That's pre-version one software in front of your production traffic. And like Diad, it's open core.
10 more repositories, a YouTuber, two tech giants, and a lot of solo builders. All giving away software you're still paying for every month. Same advice as part one. Don't install all 10. Pick the one bill that annoys you most and replace it this week. I've put all 20 repos this list and part one's into the updated free stack guide. The links, what each one replaces, and every honest catch. Comment the word cancel and I'll send it straight to you. If you missed part one, it's on screen right now. Deep dives live here on YouTube, so subscribe. And the bite-size drops are on Instagram at Hyper Automation Labs. Now, go cancel something else.