0831 | oMLX: Open-Source macOS LLM Server; Superagent, Maritime, Ulpaso

Mia: Welcome to Product Hunt Daily on Bri Radio. I'm Mia.

Milo: And I'm Milo. Today we're looking at some standout launches, from a Swift-based, open-source inference server that runs from your Mac's menu bar, to a meeting-notes app built on local markdown.

Mia: Plus an AI teleprompter that scrolls as you talk, an Overleaf-style LaTeX editor for research teams, and a graveyard for abandoned side projects.

Milo: Let's jump into the bright spots from the community this week.

Mia: oMLX just landed as an open-source, Apache 2.0 macOS-native inference server, written in Swift rather than Electron and living right in the menu bar. The maker positions it for coding-agent workflows as a drop-in backend for Claude Code, OpenClaw, and Cursor, serving LLM, vision-language, embedding, and reranker models at once through an OpenAI-compatible chat endpoint and a native Anthropic-compatible messages endpoint.

Mia: The competitive edge isn't raw speed, it's the memory model. oMLX keeps a two-tier KV cache, with hot blocks staying in RAM while cold blocks get written to SSD in safetensors format with LRU eviction. Previously seen prefixes get restored across requests and even server restarts instead of being recomputed, and the maker claims that cuts time-to-first-token on long contexts from thirty to ninety seconds down to under five.

Milo: That targets a real pain, since the maker says Ollama and LM Studio keep KV state in memory only, so a mid-session context shift invalidates and recomputes the whole cache. One community commenter confirmed it from direct experience, saying Claude Code re-read the entire conversation on every follow-up turn and that the restart-surviving cache mattered more than raw tokens per second, while another reported old context returning in milliseconds even after a restart.

Milo: On the hardware side, you need Apple Silicon M1 or later with macOS 15 and sixteen gigabytes of RAM minimum, though the maker recommends sixty-four gigabytes or more and calls an M-series Pro or Max with sixty-four gigabytes the sweet spot for daily coding work.

Milo: Superagent is a Mac app that gives your coding agent a real computer instead of keeping you in a terminal. It drives a browser on your own logged-in sessions, taps and swipes an iOS Simulator, and holds your files, a board, and routines on a timer.

Mia: So the pitch is making coding agents feel like using a good Mac app rather than a terminal. Chats sit in a sidebar you can group, they survive restarts, and each one runs on its own git worktree, so an agent's chaotic experiments don't bleed across your work. It's also private: pair your iPhone and the work follows you, encrypted end to end, with no account, no server, and no second bill. And because it's open source, it's being billed as Claude Code for the rest of us.

Mia: Maritime is taking the opposite approach to infrastructure for AI agents. Instead of a tool that runs agents on your own machine, it gives every customer's agent its own dedicated computer, starting at one dollar per agent per month.

Mia: Each agent runs in an isolated, persistent VM where it can use a browser, create files, run code, and keep its own state across sessions. So a company selling agents gets a way to hand each customer a working computing environment without building and managing thousands of VMs themselves.

Milo: Maritime handles the infrastructure, scaling, and automatic sleep and wake for you, and it's aimed squarely at companies running agents at scale, where keeping every customer's agent alive on its own machine would otherwise multiply the operational burden fast.

Mia: Ulpaso is a meeting-notes app for the Mac presented with a simple pitch: it just takes the meeting. You press record, your microphone and system audio get transcribed locally on your machine, organized by speaker, and dropped into a real, editable Markdown file in your Documents folder.

Milo: What stands out is the privacy framing. The maker contrasts it with typical meeting-notes apps that demand the recording, an account, and a credit card up front, claiming Ulpaso has no cloud, no login, no account, no telemetry, no proprietary format, and no subscription on the order of twenty dollars a month. It's described as free, MIT-licensed, and running entirely on your machine.

Mia: One community commenter said their real reason to choose it over AI notetaker bots was privacy, pointing to client calls where account numbers or other sensitive information gets read aloud and they wouldn't want that sitting in someone else's cloud transcription history. A plain Markdown file on their own disk was the only format they'd trust for that kind of call. Still, all of this is maker claims for now. One commenter asked whether it handles two people talking over each other, and the launch material contains no answer, so overlapping-speech handling remains untested and unknown.

Milo: Caplio just landed on Product Hunt as a Mac app that turns scattered screenshots and everyday images into a private, searchable library. The maker says he built it because he saved screenshots and reference images constantly but almost never found them again.

Mia: Mm—and that's the key difference from general file search. Caplio watches whichever folders you add, indexes them, and extracts text with on-device OCR, so you can later search recognized words inside images, file names, categories, and app names. Found images can be dragged into another app or copied or shared along with their extracted text, with a timeline and filters by category or date. You can also let it reorganize files on disk by date or category, previewing and restoring anything moved.

Milo: Privacy is a stated core point: OCR runs on-device, indexing stays local, no cloud AI is required, and images and extracted text are never uploaded. It's listed with Razieh Aghababa as developer, about five point six megabytes, in Productivity, requiring macOS fourteen point six or later, and it handles PNG, JPEG, HEIC, TIFF, and WebP.

Mia: And it's free with in-app purchases. Caplio Pro runs one dollar and ninety-nine cents a month, eight ninety-nine yearly, or a one-time twenty-four ninety-nine lifetime buy, with a trial giving full access first. An August nineteenth version one point one update claims better text recognition for complex layouts and improved mixed-language support.

Mia: Edge Drop is a free, open-source Windows clipboard utility that lives on the edge of the screen, invisible until you hover—then a clipboard shelf opens. You can copy text, screenshots, files, or links and drag them straight into any app. The maker says the desktop stays click-through until your cursor hits the edge, and dropping uses native Windows drop rather than simulated paste.

Milo: That native drop detail matters, because it's what makes this feel like moving a real object instead of triggering a paste command. The maker says the shelf can stack files and pin keepers, requires no shortcuts, no tray, and no account, and clipboard history never leaves the PC. It's Apache-2.0 licensed, supports Windows 10 and 11, and ships through GitHub or the Microsoft Store.

Mia: Hmm—the maker, Deepender, built it because Win+V interrupts his work. His description: copy, switch window, hunt through history, paste, then repeat all over again when he needs a screenshot in Photoshop. So this is a hover shelf on the screen edge meant to kill that loop.

Milo: In the discussion, one commenter said screenshots break their flow because a copied screenshot gets lost to the next copy, and they asked whether pinned items survive a restart or the shelf resets. Another commenter just thanked the maker and said they'd been looking for exactly this. And the maker's own open question was which copy-and-paste motion still wastes people's time.

Milo: Jensen, founder of Topview.ai, launched Motion Studio on Product Hunt—an AI workspace for building polished launch videos for apps, hardware, and new products. Its pitch is to sit between learning After Effects and hiring a motion design studio. You describe the product, the message, and the scenes you want; upload product, brand, interface, or reference images; choose a motion style; and set duration and aspect ratio.

Mia: The maker says Motion Studio combines AI video models with a motion design skill to turn that brief into a complete motion video project, and claims it's faster, simpler, and a fraction of the traditional cost. He does describe it as an early version and asks for feedback—so the workflow, speed, and cost claims are his statements, not verified results.

Milo: The community feedback focused on indie makers and small teams where the same person also handles launch content. One commenter said their team usually settles for a screen recording because motion design takes too long. Another said the workflow's starting point—the product story rather than random visuals—is exactly what launch videos need, and someone else compared it favorably against the After Effects learning curve.

Mia: Lingering on the table were a couple of genuinely open questions: whether a few screenshots and a short description are enough starting material, and whether music and sound effects are handled inside Motion Studio or have to be added separately.

Mia: Sayscroll is a browser-based AI teleprompter that scrolls as the speaker talks. Where a traditional teleprompter runs at a fixed scroll speed you adjust by hand, Sayscroll uses real-time speech recognition to fade out each spoken word and keep the script's place centered on screen. It supports sixty-plus languages, including paragraphs that mix languages, and it's aimed at YouTube creators, presenters, educators, podcasters, and public speakers.

Milo: The maker says he built it after struggling to sound natural recording a founder interview for an accelerator application—constantly adjusting scroll speed made him fall behind, get ahead, or lose his place. So the stated difference is that Sayscroll follows the voice automatically, waits while the speaker improvises, and locks back on when they return to the text, even after jumping to another section.

Mia: The source-supported set is worth laying out: there's a built-in video recorder that overlays the script on the recording or uses a split-screen layout, saves takes to the device, and nothing gets uploaded. Studio controls cover text size, column width, alignment, fullscreen, plus horizontal and vertical mirroring for beam-splitter screens. And there's a free demo accepting scripts up to seven hundred characters.

Milo: On the privacy side, the maker states the script stays on the device and only audio is streamed for transcription. Pricing is usage-based and billed monthly, with only live microphone time counted—Starter runs five dollars the first month, then ten dollars monthly for ten hours, and there's a Creator tier above that.

Mia: Murfy is an AI-native online LaTeX editor from Murple Inc., built as an Overleaf alternative for research teams. The headline claim is that users can write, review, and publish to arXiv ten times faster, and founder Shounan An from Murple says the product came out of fifteen years of writing papers accepted at AAAI, EMNLP, SIGGRAPH, and IROS in Overleaf, where formatting references, chasing co-authors, and LaTeX compile errors were the bottleneck.

Milo: The product site talks up speed and collaboration by comparison. Its table claims Murfy compiles two times faster than Overleaf, gives a thirty-second free compile window against Overleaf's ten, allows unlimited collaborators where Overleaf's free tier allows one, and bundles AI chat, comments, and track changes that Overleaf charges for, all with no per-seat pricing.

Mia: On the AI side, Murfy claims an AI paper review that works like eight readers going through a draft with three models checking each other's findings, compile-error fixes shown as an accept-or-reject diff, reference-paper verification, and Beamer slide generation. Those AI features run on credits, with a thousand free a month on the Starter plan and no credit card required.

Milo: Murple reports a hundred eighty thousand plus researchers in over a hundred twenty countries, seventy thousand plus monthly active users, and fifty to a hundred percent month-over-month growth. One tester does confirm it feels fast and handles multi-collaboration, but still hit two problems on a second editing pass: a first page labeled "Murfy Guide" that could not be removed even after replacing the report in the editor, and an issue with the AI chat response.

Mia: RIP MY BUILD is a "side project cemetery" for public postmortems. A founder can publish a permanent obituary for an abandoned project with the original link, cause of death, lesson learned, and whatever they are building next, and each memorial gets a shareable page, visitor roses, permanent plot adornments, and a weekly honors leaderboard.

Milo: The launch listing priced obituaries at twelve dollars each, but a community commenter announced they are now free permanently. The product site confirms writing and publishing an obituary is free, and community obituaries are published free after a publication review, though the review's timing and requirements are not stated. The site reports seven community obituaries so far and ten thousand seven hundred build hours documented.

Mia: Its featured editorial obituary honors Logology, built by Dagobert Renouf and Lucie Baratte, dated 2018 to 2023, with the epitaph: "We spent five years perfecting our product. We shipped too slowly to finish the brand." The weekly honors count roses and adornments on each memorial, and the register resets every Monday at midnight UTC. There's also a commonplace book ranking the most "stolen" lessons with repeat clicks filtered, where the top lesson, from Idento, is "A clever engine is not the product," with advice to start on one narrow naming problem.

Milo: In Product Hunt discussion, one commenter admitted they kept polishing the product before ever putting up a landing page and should have shipped a one-page waitlist in week one, while another described landing their first paying customer.

Mia: Before we go, the big takeaway: oMLX turns your Mac into a local LLM server that answers in seconds, not minutes — and Superagent gives your coding agent a computer, not a terminal.

Milo: Two very different ways to put your own machine to work. Thanks for listening, and bottom's up.

Mia: Cheers, and see you next time.