0822 | Vercel's tiny fx, Base Compute Local, Plow Latch, Antigravity IDE

Mia: Welcome back to ProductHunt Daily on Bri Radio. I'm Mia, and I'm joined by Milo. Today we've got a fresh lineup of launches, from a tiny open-source coding agent out of Vercel to AI that runs fully offline on your own Mac. Milo, what's catching your eye first?

Milo: Honestly, the Google news is huge - Antigravity, their agentic coding platform, is coming to Visual Studio through new extensions announced back in August. But there's also Local, a free Mac app from Base Compute that keeps AI entirely on your machine with no cloud and no account required.

Mia: And that's just the start. We've got tools giving AI agents scoped access to your Mac, a personal AGI pitch for your PC, and even a Windows desktop companion built as part of the operating system itself. Plus Vercel's newest coding agent that's designed to get out of your way.

Milo: Right, from tiny and invisible coding agents to OCR utilities that pull text off any part of your screen. Lots to dig into, so let's get going.

Mia: There's a new free Mac app called Local from Base Compute, and its whole pitch is running AI privately on your own machine—no cloud, no account, no per-token cost. The key claim is zero-friction setup: it auto-tunes to your hardware and recommends models your machine can actually run, so you skip the manual model configuration that usually comes with on-device AI. Base Compute says that on the same model and the same Mac, throughput runs up to five point four times faster than the engines other local apps ship with, because its BaseRT inference engine is compiled for your specific chip on first launch rather than shipped as a generic build. A community member pointed to BaseRT research to back that up, and one commenter explicitly asked to verify the auto-tuning.

Milo: The app only needs internet to download the app and a model—after that it runs offline, and the maker says no telemetry is collected and everything lives in a database on the Mac. What stands out is the range of what it does locally: chatting with files and summarising PDFs, plus a coding mode that reads, edits, and runs code in a local folder without ever uploading the codebase. There's also on-device meeting transcription with speaker labels, summaries, and decision logs, and memory you can mark sensitive so it never leaves the machine.

Mia: It even tracks the tokens generated and the estimated cloud cost you saved by running on-device. And there's an office mode where every laptop in a room can connect to one faster machine that's running the big model. It supports open models like Qwen, Llama, Gemma, Mistral, Phi, DeepSeek, and Whisper, and if you want, you can plug in frontier models with your own API keys for OpenAI, Anthropic, or OpenRouter.

Milo: Plow Latch is a Mac app for giving AI agents scoped access to the machine. The maker describes it as a way to safely give an agent real control of the Mac while keeping data local, so pretty much anything you can do with your Mac, you can do with your AI. It works with AI you already use, including Claude.ai and Codex, lets an agent drive the browser and a command-line interface, can run fully locally, and uses an adversarial LLM to gatekeep data in case an agent goes rogue. The maker says Latch makes sure the agent only has the tools and credentials it needs, and reports using it with Claude.ai to pay bills, organize a calendar, and order groceries and food.

Mia: Right—and the scoping concern is real. If several agent sessions on one machine share a browser, one session's routine page read could pick up another session's single-use OAuth code from a localhost callback page. So the open question is whether a single-use secret meant for one session can leak to another when they share one browser state. Someone also asked whether similar phone functionality exists, noting that Apple used to have protections against apps like this, and another commenter asked the agent for the password to a DoorDash account.

Milo: The maker went further than a typical launch: they opened their own Mac for a live demo, so anyone could leave a Product Hunt comment and watch the Mac respond on a Twitch livestream. Commenters tested it by ordering food and drink with specific delivery locations, and one asked the agent to build a local HTML page styled like an old eight-bit game. The app is available for download at plow.co/latch.

Mia: ShogunAI is a macOS app available today, launched under the tagline "built to finish real work." Its maker defines "personal AGI" not as human-level intelligence but as general across your work rather than narrow to a single task—one agent that spans the whole day, holds the state of it, and acts on that state. Your work state stays on your own machine rather than in a cloud account. It builds a single record of people, projects, promises, and open items from your day, keeps the evidence behind every record, reads automatically, and waits for approval before anything gets sent to another person.

Milo: Capabilities include inline drafting at the cursor via the Option key written locally, meeting minutes with summary, decisions, and commitments where audio is never written to disk, two-way live translation that keeps only text, and recall across mail, chat, docs, and calendar with source and time attached to every hit. Autonomy is tiered: reversible work runs on its own, drafts wait for approval, and anything leaving the Mac stops for approval.

Mia: It runs either on the assistant plan you already pay for or on your own API key, and every action leaves a record of what ran, what evidence it used, and what left the device. The maker contrasts it with Hermes, which learns from its own sessions—ShogunAI instead passively holds the state of the day and serves it back.

Milo: Project SKY is a Windows desktop AI companion making the opposite argument—that AI shouldn't be trapped inside a browser tab. The builder says he built it to bridge the gap between AI and the actual operating system, positioning it as an integrated OS layer with spatial screen perception, conversational voice, long-term ambient memory, screen analysis, automated desktop workflows, and voice-driven task management, meant to understand what you're working on across applications without breaking focus.

Mia: His claims are that it sees the screen in real time, converses over voice with zero latency, and retains persistent long-term memory across sessions without repetitive prompting. He describes the design as evolving from a standard chat widget into an integrated OS layer, with circle-to-search screen selection and native desktop orchestration coming straight from the goal of making interaction feel effortless. Worth noting that all of this is the maker's own description—the launch material includes no independent verification or third-party test results.

Mia: PixelRead AI OCR is the macOS answer to text trapped inside screenshots, videos, PDFs, and apps. One shortcut, Command-Shift-2, plus a drag over any region fires instant OCR through Apple Vision, and the captured text can be copied, translated on-device beside the original, read aloud with language-aware voices, or handed to Apple Intelligence to summarize, rewrite, pull out names, dates, numbers, links, decisions, and action items, and answer questions grounded only in the recognized text.

Milo: The privacy angle holds because the maker says OCR, translation, and AI processing all stay on the Mac with nothing sent to a server, and translation runs through Apple's on-device framework. On macOS 26 the translation also auto-detects the source language. The catch is that translation and the Apple Intelligence features require macOS 26 and supported hardware, while base OCR needs macOS 15.2 or later and runs natively on both Apple Silicon and Intel.

Mia: And it's free, with no purchase needed — the current version is 1.1.2, and it was 1.1.0 that dropped the trial and lock screen. The maker built it because existing OCR utilities usually stop at copying, so the goal was one fast shortcut that could translate, speak, summarize, and answer questions without sending captured text anywhere, and they're specifically seeking feedback on the capture workflow and which text actions people use most. Notably, one Product Hunt commenter said the translation feature looks useful and that they'd mainly use it for that.

Mia: Vercel's fx is an open-source coding agent pitched as tiny and out of the way. Written in Zig, it ships as a native binary of roughly six megabytes — version 0.0.4 lists at 6.39 mebibytes — starts almost instantly, and keeps memory and context overhead low. One commenter made the point that its small system prompt and tool surface leave more of the model's context window for the actual work.

Milo: It's also model agnostic, so it works with local or cloud models, and it supports WebAssembly — in fact the product site runs the full fx command line as WebAssembly compiled with the Zig toolchain, with networking delegated to the browser's fetch. It can be extended with skills, plugins, and MCPs, or embedded into your own agent infrastructure.

Mia: The status is the real caveat though. It sits at version 0.0.4 with experimental status, and the maker warns to use it at your own risk with frequent changes coming, so long-term stability and interface continuity remain untested and there's no promise of a stable release. The community praise — one commenter said it starts almost instantly, uses very little memory, and feels more like a Unix tool than another full IDE in the terminal, while others called it beautiful and praised the open-source angle — those remain attributed impressions, not verified results.

Mia: Google brought its agentic coding platform into four editors with Antigravity IDE Extensions, announced August 20, 2026. Visual Studio Code, Visual Studio, JetBrains, and Zed now keep agent conversations and shared context while developers review inline diffs, inspect plans, debug code, and hand off multi-step tasks without leaving the editor. Google frames them as lightweight integrations, not attempts to turn existing editors into another Antigravity, and the standalone Antigravity 2.0 app stays the flagship for end-to-end multi-agent coordination.

Milo: The rollout is uneven by editor. Visual Studio Code gets a lightweight extension from the marketplace, searched for and signed in to from the Activity Bar. Visual Studio installs through its marketplace or extension manager and is currently in Preview. JetBrains covers the IntelliJ-based suite including IntelliJ IDEA, PyCharm, WebStorm, GoLand, CLion, and Rider starting with version 2026.2.1 with one-click install, and Zed integrates the same way.

Mia: Access tiers differ too. Individuals sign in with a Google account on any Antigravity plan, including the free tier, while enterprise teams use Gemini Enterprise or Google Cloud project credentials, with enterprise support for JetBrains and Zed currently in Preview. Sessions run under the Google Cloud Terms of Service, and Google states the data is never used to train foundation models.

Mia: OneCLI is an open-source, self-hostable agent harness for teams that gives every employee a secured, sandboxed assistant agent usable in Slack and on the web. Co-founder Jonathan says he built it with Guy because an AI agent needs access to real systems like GitHub, Gmail, a CRM, or cloud, but handing an agent real passwords or keys is risky since keys can leak. The core design is that each employee's agent never holds a real credential — it sees only a placeholder, and the OneCLI gateway injects the actual secret at the network layer for each request, but only after the request is approved.

Milo: For sensitive actions like sending an email or deleting a ticket, the agent asks a human first. The site illustrates this with an agent checking 1,204 Stripe charges against company books and stopping at an $840 refund because payments over $500 require human approval. The claimed guardrails are hard limits outside the model against deleting a repository, sending a payment, or wiping a customer record, plus slowing or pausing an agent that repeats itself or works far faster than a person.

Mia: Jonathan says he previously built zero trust network access, or ZTNA, at Axis Security, which was acquired by HPE, and that Guy was the first employee at Argon, acquired by Aqua Security. OneCLI is Y Combinator-backed, fully open source with more than 350,000 downloads, and offers a free tier with no card required, with teams able to self-host.

Mia: Actx0 just hit Product Hunt, pitched by its makers as managed memory infrastructure for AI agents. The argument is that agents forget everything when a session ends, so developers keep feeding old data back into prompts and pay a heavy token tax. Actx0 is meant to be a drop-in memory layer that extracts what matters, retrieves it in milliseconds, and carries it across sessions, agents, and apps, with no vector-store babysitting and no bloated prompts.

Milo: One commenter who runs agents for multiple customers pushed back on that babysitting point, saying it eats more engineering time than anything else in their stack and zeroing in on tenant separation. They asked whether isolation is enforced at the storage layer itself, with separate namespaces or keys per customer, or whether it relies on how memories get tagged and queried, arguing a memory leak across customers is a much worse bug than an agent just forgetting.

Mia: That question stays open. Another commenter asked whether Actx0 works with OpenClaw, and a third said the SDK feels intuitive, supports multiple frameworks, and asked whether the cloud is fully managed. The makers call the release day one of a massive roadmap, saying it's not yet a flawless final product and that they're shipping fast. It's free to use right now, with paid plans coming soon.

Mia: Router by Ramp is a separate launch with Ramp pitching a single API endpoint that routes every request to the lowest-cost model still meeting a performance threshold, aimed at teams scaling multi-agent systems whose inference bills balloon while they juggle separate API integrations, rate limits, and fallback strategies. The launch comment claims one endpoint covers OpenAI, Anthropic, SpaceXAI, and open-source models through a single API key with no codebase rewrite, with auto-routing matching request complexity to the right model.

Milo: Ramp also claims built-in spend visibility that pairs inference routing with its financial engine to map token usage back to teams and budgets, instant setup, a zero-cost routing layer through 2026, twenty-six dollars in model credits, and an average forty percent saving on inference spend. These are all maker and community claims, not verified results.

Mia: Community reaction turns on the decision itself. One commenter said they burn energy second-guessing which model to use and find letting that decision settle automatically appealing, with clearer visibility as a bonus. Another said LLM costs get messy quickly and that balancing price and performance is worth solving. One asked how Router differs from OpenRouter and from Cortecs, but the discussion never answers that.

Mia: So today we looked at two very different takes on AI tooling — Vercel's tiny open-source coding agent that stays out of your way, and a free Mac app that runs models privately on your own hardware.

Milo: Both built around a similar idea, really: less overhead, more of your control. Thanks for listening — we'll catch you next time.