0718 | Kimi K3, Pebbles Ai, and Unabyss for Claude

Mia: Welcome to ProductHunt Daily, a podcast from Bri. I'm Mia.

Milo: And I'm Milo. Today we've got some great headlines lined up for you — from Kimi's new open-source model to a fresh tool called Unabyss built for Claude users.

Mia: We'll also look at Pebbles Ai, Aye, and several other launches making noise on ProductHunt right now.

Milo: Let's get into it.

Mia: Pebbles AI launched on Product Hunt this week, introduced by Chief AI Officer Dmytro Antoniuk. It is a Go-To-Market Operating System built specifically for B2B revenue teams.

Milo: And the core idea here is consolidation. Pebbles AI pulls strategy, lead generation, outreach, sales, and shared company knowledge into a single AI-powered workspace. The pitch is that you stop juggling a pile of disconnected GTM tools.

Mia: Right. Under the hood it uses something they call neurosymbolic AI, and it is trained on a customer's own business to plan campaigns, personalize outreach, generate qualified leads, and actually execute the workflows.

Milo: Antoniuk was pretty specific about where this is aimed. He highlighted that chaotic early stage when a team is trying to land those first customers. You know the moment: you are doing market research, hunting for target leads, writing cold outreach, managing follow-ups, all at once.

Mia: Exactly. And that is the mess Pebbles AI wants to clean up. One workspace that handles the whole cycle from strategy to close.

Milo: Switching over to models, Moonshot AI has introduced Kimi K3, and they are calling it the world's first open three-trillion-parameter-class model.

Mia: Three trillion parameters is the class, but the actual count is 2.8 trillion. It is a mixture-of-experts model with a one-million-token context window and native multimodality, built for long-horizon coding and knowledge work.

Milo: Now the developers are upfront that overall performance still trails the most powerful proprietary models. They specifically name Claude Fable 5 and GPT 5.6 Sol as the ones they are chasing.

Mia: But Kimi K3 did demonstrate frontier-level performance across their evaluation suite and consistently outperformed every other model they tested.

Milo: And the reported capabilities are worth noting. Working across large code repositories, using vision while iterating on games and frontend tasks, and sustaining longer agent runs with limited supervision.

Mia: It is accessible today through Kimi.com, Kimi Work, and Kimi Code. So if you want to kick the tires on an open model at this scale, the door is open.

Mia: Also on Product Hunt, a new browser called Aye from builder Zhonglin Liu. It is a Chromium-based AI browser for macOS and Windows, and the framing is interesting. It acts like a teachable AI intern.

Milo: A teachable intern that reads the visible page, plans its steps, and then works through normal browser actions. Clicking, typing, scrolling, switching tabs, checking results. All the things you would do, just handled by the browser.

Mia: And the use cases stack up. It can summarize and translate pages, research across multiple tabs, draft content, and automate repeatable workflows. Those recurring tasks can even be turned into reusable skills.

Milo: There is a trust and control angle here too. Aye offers separate accounts with profiles, and it keeps users in the loop with reviewable progress and approval gates for sensitive steps.

Mia: The problem Liu is solving is familiar. A lot of browser work starts with a simple goal and then turns into this long chain of reading, comparing, clicking, and rewriting. Aye carries that work forward through visible, reviewable actions on real websites.

Milo: And one more from Product Hunt. Unabyss is now available for Claude, and it is built around a really practical problem. Giving different large language models and applications a shared memory.

Mia: So Claude can receive memories from other AI agents and from everyday apps. Email, Google Drive, GitHub, Notion, meeting recorders, and more than twenty additional services.

Milo: And it works in the other direction too. Unabyss saves new memories so that GPT and Cursor stay in sync with the same context. The system is described as private, portable, and a real memory that follows the user.

Mia: It operates inside Claude through a Claude-first, MCP-first approach. And once you connect the MCP, you do not need a browser anymore. That is a nice detail.

Milo: Co-founder Philip noted that the original launch back in May won the number one product of the day award on Product Hunt. This relaunch followed user feedback, and they rebuilt the product around the idea that context should live where users actually work.

Mia: Basedash is rolling out a feature called Suggestions, and the idea is straightforward: instead of waiting for you to ask a question, the AI data analyst comes to you with proposals.

Milo: Right. So the system looks across your connected data sources, your past chat history, and the dashboards you already have set up. Then it surfaces tailored prompts, dashboard ideas, and automation candidates.

Mia: And it is not one-size-fits-all. The suggestions are tuned for specific roles: growth teams, finance, operations. The launch was announced by Max from Basedash.

Milo: What is interesting is how the loop works. You click a suggestion, that kicks off the actual analysis, and once it is complete, the item gets replaced with a fresh recommendation. So the feed keeps turning over.

Mia: It flips the interaction model, doesn't it? Instead of you having to know what to ask, the tool is constantly surfacing what it thinks is worth investigating based on what it already knows about your data.

Milo: Exactly. And for teams that live inside their data but do not always have the time to formulate the right query, that could change how they start their day.

Mia: Switching gears, there is a new tool called Scribble Party, and it is a local-first whiteboard studio built with teachers and creators in mind.

Milo: It runs entirely in the browser on an infinite canvas. You can place a live camera bubble of yourself, with real-time background cutout, right on the board.

Mia: So you are drawing or laying out content, and there is a little video of you talking, sitting directly on the canvas. And you can record the whole thing, canvas plus your voice.

Milo: Here is the part that stood out to me: recordings are saved directly to your device as you go. So if your browser tab crashes mid-recording, you do not lose the take. The file is already written to disk.

Mia: That is a meaningful detail. And it works offline as an installed app. No signup, no installation, nothing ever uploaded.

Milo: One user who tested it noted the background cutout held up well during movement, which is not always the case with browser-based tools. They also appreciated seeing the file actually write to disk while recording.

Mia: Another user raised an interesting point though: the no-upload privacy model is great for peace of mind, but it does raise questions about discoverability. Your work stays local, so how do people find what you have made?

Milo: It is the classic tradeoff. Total privacy and local control on one side, sharing and reach on the other. Scribble Party picks the first one.

Mia: Let's talk about Timely. This is a tool for a very specific, very common pain point: pulling calendar availability across timezones.

Milo: The pitch is dead simple. One keystroke pulls availability from your connected Outlook and Google calendars, and it automatically handles the timezone conversion. No opening a calendar, no manual maths.

Mia: And it covers multiple executives, which is where this gets interesting. The maker built it to replace a very specific chore.

Milo: Right. The manual cycle is: open Outlook, scan for open slots, convert to the recipient's timezone in your head or in another tool, and then type out a list of options. If you are doing that for multiple people, it compounds fast.

Mia: And it is not just executive assistants. The maker says it is also for people who self-schedule across timezones and for personal assistants booking on behalf of principals.

Milo: Everything runs locally on your machine, and your data stays private. The positioning is interesting too: it is pitched as a dead-simple alternative to booking links, for situations where clients do not want to do the scheduling legwork themselves.

Mia: So instead of sending someone a Calendly link and saying 'you figure it out', you can fire off three time slots that genuinely work across both calendars and both timezones, in seconds.

Mia: Finally, PixyCAD. This is a native 3D CAD app built for Mac and iPad, aimed at makers, designers, and 3D-printing enthusiasts.

Milo: What sets it apart is the engine underneath. It runs on Parasolid, which is a professional geometry foundation. So you are getting serious solid-modelling capability, but in a package built from the ground up for Apple platforms.

Mia: The team, including Andrea, created it because they felt Mac and iPad users did not have an approachable native modelling experience that still gives them professional-grade tools.

Milo: Workflows cover the full chain: sketching through to solid modelling. So you can start with a rough shape and build it into a manufacturable part.

Mia: And they are launching on Product Hunt specifically to gather feedback from people who care about CAD, 3D printing, and better creative tools on Apple platforms. So it sounds like they are treating this as a conversation starter, not just a launch-and-done.

Milo: That makes sense for a CAD tool. The community that uses this kind of software has strong opinions, and getting that feedback loop going early is smart.

Mia: And for anyone doing 3D printing from a Mac or iPad, having a native app with that Parasolid foundation is a pretty compelling combination.

Mia: So Pocket Screen is a new Mac app from indie developer Masaki at Toybird Labs, and it does one thing really well.

Milo: It takes whatever window is frontmost on your Mac and turns it into a compact, always-on-top floating view. Think picture-in-picture, but for any app.

Mia: Right, so you could keep a document visible while you're writing in another app. Or a chat window, a reference video, anything you'd normally have to keep switching back to.

Milo: And the mini window isn't locked in place. You can move it around, resize it, make it transparent, or even set it to click-through mode so it doesn't intercept your mouse when you're working behind it.

Mia: That click-through mode is a nice touch. You get the visual reference but the window stays out of your way entirely.

Milo: Control comes from the menu bar or a global shortcut, so you're not hunting through panels to adjust it.

Mia: And here's the privacy piece: all processing stays on your Mac. Nothing leaves the device.

Milo: That's worth noting because a lot of window-management and screen-capture tools these days push things to the cloud. Pocket Screen keeps everything local.

Mia: It's a focused little utility from an indie developer, Masaki at Toybird Labs. The whole idea is reducing window switching and keeping your reference material visible without breaking your flow.

Milo: And because it works on whatever's frontmost, it's pretty flexible. You're not locked into a specific set of supported apps. Anything you can bring to the front, you can float.

Mia: So if you spend a lot of time bouncing between windows just to glance at something, Pocket Screen gives you that persistent visual anchor. Move it, resize it, fade it, or let clicks pass through, all controlled from the menu bar or a shortcut, and everything stays on your Mac.

Mia: And that's our show for today. We had a packed lineup — nine tools, platforms, and experiments worth watching.

Milo: Let's do a quick loop back through. Pebbles AI turned a single image into an interactive 3D world — instant scene reconstruction in the browser.

Mia: Then there's Kimi K3, Moonshot AI's open-source reasoning model. It's already available on Hugging Face and claims performance competitive with models much larger than itself.

Milo: Aye caught our attention too — it's an AI coding agent that runs entirely in the browser with WebContainer tech. No setup, no cloud dependency.

Mia: We also looked at Unabyss for Claude, which extends the model's context window to 100,000 characters and strips the refusal layer — that one's a deliberate provocation worth watching closely.

Milo: Basedash Suggestions, Scribble Party, Timely, PixyCAD, and Pocket Screen rounded out the rest — each doing something small and sharp, from AI-powered UI iteration to collaborative drawing to real-time CAD generation.

Mia: Great range this episode. We'll have links to everything we covered in the show notes. I'm [Host A] —

Milo: And I'm [Host B]. Thanks for spending this time with us. We'll catch you on the next one.