0722 | Terminal, Browser, Claude Code & AI Diff Reviews

Mia: Welcome to ProductHunt Daily, the Bri podcast that brings you the most interesting new launches every day.

Milo: I'm Milo.

Mia: And I'm Mia. Today we've got a packed show — from a macOS cockpit that combines a terminal, browser, and Claude Code into one app, to a code review studio built for AI-generated diffs, and a tool that helps teams optimize for answer engines instead of search engines.

Milo: Plus we'll look at a hardware-isolated sandbox for AI agents that provisions in about 30 milliseconds, a voice AI agent studio, an AI checkout agent that's been in the works for over a year, and a clever menu bar utility that renames your screenshots from their actual content.

Mia: First up — tterm. This is a macOS app that the creator describes as a cockpit. It combines a terminal, a real Chromium browser, and Claude Code into a single application.

Milo: Right. And the layout is interesting. Each project shows up as a row with three pieces: an explorer on one side, a Claude pane, and a file viewer. What's notable is what's missing — there's no built-in editor by design.

Mia: The idea is you review diffs hunk by hunk and commit. The creator said they were living across three windows — a terminal, a browser, and an editor they barely typed in anymore — and wanted all of that folded into one place.

Milo: And the browser piece isn't some stripped-down web view. It's a real Chromium browser that carries over your Chrome bookmarks and cookies. So all your saved sessions and logins are right there inside the app.

Mia: Here's a wild detail: tterm can rebuild itself. You ask the Claude pane for a feature, and the running app hot-reloads with that feature baked in. You're essentially coding the tool you're using, from inside the tool.

Milo: No account required, no telemetry, and it's free for hobbyists on Apple silicon. That's the full picture on tterm.

Mia: Let's move to Diffsmith. This is a code review studio built specifically for the AI coding agent workflow. It works with local uncommitted git changes from tools like Claude Code, Cursor, Codex, and Copilot.

Milo: So here's the flow: you open a repository, read the diff, and click any line to leave a comment that's anchored to that exact spot. Then you hand those notes back to the agent. It's designed for that back-and-forth between you and the AI.

Mia: And it supports MCP — which means you can actually view the agent's replies to your comments directly at the related line of code. So the conversation about the code is attached to the code itself.

Milo: The tool is a hundred percent offline. And it's a one-time purchase — you pay once, and you get lifetime access. No subscription.

Mia: One piece of community feedback worth noting: users have asked for a side-by-side mode that shows the original file alongside the diff. Right now, if you want to recall the surrounding context before commenting on a single line, you have to flip back to your editor. That's something people want addressed.

Milo: Now, Manifest. This comes from a solo founder named Max, who runs a company called Omfang AB. He built it to solve a very specific pain point.

Mia: The problem he kept hitting: every time he tried to get an AI agent to reliably interact with a webpage, he ended up hand-writing selectors that broke the moment the site's DOM changed. He looked around and saw that browser automation tools handle providing a browser, and content extraction tools handle providing the content, but nothing addressed the core need — telling an AI agent what is clickable, fillable, and submittable, and how those actions depend on each other.

Milo: So Manifest turns any webpage into a structured JSON map of available actions, using a single API call with no fragile selectors. Every action includes resolved CSS and role locators. And here's the clever part — it encodes dependencies between elements through a requires field.

Mia: Right, so for example, it might note that the agent has to select a plan before a specific button becomes clickable. The agent gets that relationship baked into the map.

Milo: The project ships with a Python SDK, LangChain support, and an MCP server. It's targeted at anyone building browser agents. That's Manifest.

Mia: Finally, Phantomstory. This is a new product from The Letter Company, launched by founder Mathew. It's aimed at a space called AEO — answer engine optimization.

Milo: The pitch is that teams need to be discoverable not just by search engines, but by large language models like ChatGPT and Claude. And the approach here is different from traditional SEO.

Mia: Phantomstory creates third-party blogs on fresh domains in under five minutes, using just two clicks. It sets up content hubs that incrementally publish articles written specifically so that LLMs can discover, understand, and recommend a company.

Milo: The reasoning behind this: LLMs trust third-party sources more than a brand's own self-ranking content. So instead of trying to rank your own site for an answer engine, you create an ecosystem of independent-looking content that the models pick up.

Mia: Phantomstory frames AEO as an actionable strategy rather than a monitoring problem. You're not just watching what models say about you — you're shaping it. And that's our four for today.

Mia: CreateOS just launched a hardware-isolated sandbox for AI agents, and the speed number jumps out: about 30 milliseconds at p90 to provision a new environment.

Milo: Yeah, 30 milliseconds at the 90th percentile. CTO and co-founder Pratik explains that every AI agent gets its own guest-kernel environment, so they're not sharing kernel space at all.

Mia: And the egress story is where the isolation gets real. Policy is enforced inside the kernel through eBPF, so even if code running inside the sandbox is fully compromised, it still cannot bypass those rules.

Milo: Right, the kernel itself is the gatekeeper. That eBPF egress control supports allowlists by host, by IP address, or by CIDR range, so teams get fine-grained control over what the agent can reach.

Mia: Then there's the snapshotting. Teams can snapshot a running sandbox and fork it in milliseconds. They can also establish encrypted peer-to-peer mesh networking with a VPN link back to a local machine.

Milo: And storage too. You can mount any S3 service as a shared disk across sandboxes, which is interesting for multi-agent workloads that need a common filesystem.

Mia: On the developer experience side, there's a Stripe CLI integration with createos slash project that provisions credentials directly into a project's dot env file.

Milo: And for teams evaluating whether this fits their workflows, the package includes a CLI, an SDK, over fifty worked examples, Claude plugins, and compute-SDK integration. So a lot of on-ramps to try.

Milo: Staying with agent tooling, Bolna launched something called Agent Studio, and the pitch is that you can build and deploy a production-grade voice AI agent without doing any prompt engineering at all.

Mia: The workflow is pretty simple. You upload a document or answer a set of guided questions, and then the platform assembles a call-ready agent from production-tested modules. The goal is sign-up to calling on the same day.

Milo: And the team is upfront about why they built this. They say the starting point has been intimidating, with hours needed just to get context, guardrails, and conversation flows right, and that failures still slip through at deployment.

Mia: So the idea is to take that upfront friction away by modularizing the pieces that are already proven in production, rather than asking every team to reinvent the prompt from scratch.

Milo: Now here's a different kind of agent. CartAI is an AI agent that handles checkout, and its developer release is shipping now after more than a year of building, according to founder Manil.

Mia: The gap they kept running into is an interesting one. Browser automation can navigate the web but cannot pay, and payment APIs can move money but cannot navigate. CartAI is designed to close that gap.

Milo: So the API lets an agent reach the checkout button and then complete the purchase on any live merchant surface, with no merchant-side work required. And they're cooperating with bot detection instead of trying to evade it.

Mia: That's a notable design choice, working with bot detection rather than around it. The offering is a developer-first API with four products.

Milo: Let's walk through them. Catalog handles search and live pricing across merchants. Checkouts clears and tracks orders all the way to confirmation.

Mia: Then Payments moves PCI off the developer's stack entirely by using Visa and Mastercard agent rails. And Monetization provides commissions on every agent sale.

Milo: The next signal here is straightforward. The developer release is available now, and the team is inviting integration.

Mia: Let's shift to Mac utilities. Two apps caught our eye. First, Tidy handles three desktop chores from the menu bar.

Milo: It archives new screenshots, renames them from their content using on-device OCR, ejects mounted DMG files, trashes the installer, and sorts new downloads by type. All from the menu bar.

Mia: And the sorting rules are written in plain English. Apple Intelligence builds them on-device, so nothing leaves the Mac. It only touches new files and it never deletes anything.

Milo: Pricing is nine dollars once, no subscription. Then there's MeetIsland, which turns the MacBook notch into a Dynamic Island for meetings.

Mia: So your next meeting lives around the notch and expands with a live countdown and a one-click Join button for Meet or Zoom right before it starts. It requires zero setup and reads calendars already synced to macOS.

Milo: And calendar data never leaves the Mac. It also works on Macs without a notch by using a virtual island, so it's not limited to the latest hardware.

Mia: MeetIsland is six ninety-nine once with a fourteen-day free trial. Two small utilities, two one-time prices, both keeping data on-device.

Mia: Skim is a free, open-source email client for Windows, MIT-licensed, and built with a native Rust core on Tauri 2 — no Electron anywhere. The installer is roughly 5 megabytes, cold start is sub-second, and full-text search is instant.

Milo: It is local-first: IMAP syncs offline into a local SQLite database, there is zero telemetry, and no subscription. You can bring your own Anthropic or OpenRouter key, and the AI drafts replies in your voice and answers questions across your entire mailbox.

Mia: The interface is keyboard-first, with a Ctrl+K command palette, and the creator explicitly notes there is no calendar and no rules engine.

Milo: Also on the free and open-source front — OpenChatCut, which is AGPL-licensed and local-first. This is an AI video editor where agents work on a real multitrack timeline, not a one-shot black box.

Mia: It ships with a built-in agent and also supports external agents through MCP, including Codex or Claude Code. Exports come with clips, transitions, captions, and effects.

Milo: The creators note it is free and independent of ChatCut, and they warn that macOS may flag the unsigned app in this early build.

Mia: That's our show for today. Nine tools and launches, all pushing AI in different directions – from the terminal to the checkout page.

Milo: Yeah. tterm putting a browser and Claude Code inside one macOS app, Diffsmith making local diff review work across Claude Code and Cursor, Manifest bridging AI-to-browser interactions, Phantomstory going after answer engine optimization...

Mia: CreateOS with that hardware-isolated sandbox – around 30 milliseconds at p90 for provisioning. Bolna shipping Agent Studio for voice agents without prompt engineering. CartAI handling actual checkout as an AI agent, now in developer release.

Milo: And the smaller but really practical ones: Tidy handling screenshots, renaming, and clipboard from the Mac menu bar, plus Skim – that open-source Windows email client built on Rust and Tauri 2, zero Electron.

Mia: It's striking how many of these are solving real friction points – things developers and teams actually hit every day.

Milo: That's the thread, I think. Less hype, more 'I built this because it hurt.' Thanks for listening – we'll catch you on the next one.