Open Source Radar — August 18: video, pentests & Mac inference
The coding-agent parade took a day off. Today's daily board is a 107k-star video factory still shipping, an AI pentester that validates its own findings, a Mac inference server built for real coding work, and a memory layer that survives switching vendors.
MoneyPrinterTurbo (Python, ~107k stars) — Give it a topic or keyword and it writes the script, pulls footage, burns in subtitles and music, and hands you an HD short. It has been around since 2024 and still landed on today's board because it is still shipping: a new release last week, a fresh video-generation backend today, and four ways to run it — agent, WebUI, API, or CLI. It talks to a long list of model providers, uses stock footage or your own files, and can publish straight to TikTok, Instagram, and YouTube Shorts. If you want a local short-video pipeline instead of another hosted clip toy, this is the one people keep starring.
Strix (Python, ~54.7k stars) — Autonomous AI pentesters that run your app the way a human attacker would, then come back with working proofs of concept instead of a scanner's maybe-list. A team of agents handles recon, exploitation, and validation, and it can draft a fix and a report when it is done. It also dropped a skill pack so Claude Code, Cursor, or Codex can kick off a scan, and it wires into CI so a pull request can fail closed. Use it when you want a red team that keeps pace with deploys — and treat the findings like any other pentest: verify before you merge.
oMLX (Python, ~19.2k stars) — An Apple Silicon inference server you run from the menu bar: an OpenAI-compatible API, continuous batching, and a two-tier cache that keeps past context on SSD so it survives a restart instead of being recomputed. That last part is why it is hot — it makes local models usable with tools like Claude Code, where long context is the whole job. You can pin everyday models in memory, auto-evict the rest, and serve text, vision, embeddings, and rerankers from one process. If your Mac is already the inference box, this is the control plane that finally feels like a product.
ai-memory (Rust, ~2.4k stars) — Long-term memory for coding agents, built as a shared markdown wiki compiled from sanitized session hooks. Quit Claude Code mid-task, open Codex in the same directory hours later, and the next agent gets a "where you left off" block instead of a blank chat. There is no vector database to babysit — the wiki is a git repo you can search, open in a notes app, or back up like any other folder. Use it when your work spans more than one agent vendor and you are tired of re-explaining the architecture every morning.
Worth watching this week: the board is voting for infrastructure — inference, memory, and the jobs agents still do badly — not another chat wrapper.
If you could only install one of these this week — the video factory, the pentester, or the Mac server — which one? Tell us in the comments.
Sources: MoneyPrinterTurbo (GitHub) · Strix (GitHub) · oMLX (GitHub) · ai-memory (GitHub)
