0829 | Almanac, Caddi's Screenshare Agents, screenpipe, Aramb

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

Milo: And I'm Milo. Today we're digging into the Almanac, plus agents built from narrated screenshares, screenpipe, and Aramb.

Mia: We'll also get to a haptic control deck for iPhones and iPads, a local-first prompt engineering workbench, and a coding search index from Firecrawl.

Milo: Plus MacBook notch utilities, a native Mac screen recorder, Microduck, the Spline V2 3D editor rebuild, and the Gemini Omni flash model. Let's get into it.

Mia: We're starting with two launches that drop AI agents into tools teams already live in. First is Almanac, an agent that knows your company — you connect your accounts, it builds a self-updating brain from your work, and it lives right in Slack and iMessage to actually get tasks done. One of the three founders, Kushagra, shared it on Product Hunt, and the key selling point is that it has its own computer, so it keeps working even after you close your laptop.

Milo: Second, OpenTag is basically an AI coworker that sits right in Slack and Teams, carries full context on the company, and takes real work off the team's plate. You tag it in a channel, hand it a task, and the result lands back in the thread where the whole team can see it. For the launch they're offering fifty percent off for two months.

Mia: Caddi takes a different angle from Almanac — it builds agents by watching you work. Show it a task once on your screen, and it turns that narrated screenshare into a production agent that runs across your real tools. Maker Jason says it's built for law firms, RIAs, and accounting firms, where the work isn't hard, it's just endless. Take signed contracts: they get downloaded, renamed, filed, and logged in the CRM, forty times a month.

Mia: The positioning is against RPA and workflow builders, which typically need scoping, developers, and step-by-step setup. Instead it reads your stack, ranks repeated work by impact, and teaches through a screenshare and conversation. When something is ambiguous, it stops and asks — with options and a confirm button — rather than guessing. It combines AI reasoning where judgment is needed with deterministic code where it isn't, runs in the cloud, and can execute thousands of actions at a time.

Milo: Commenters did find the screenshare approach more natural than building workflow diagrams by hand. But they flagged edge cases still open: small process changes, exceptions, workflows that stretch across three or four apps, website layout changes, processes with no API, healthcare use cases, and editing just part of an existing workflow. The website claims SOC 2 Type II and over a hundred fifty integrations, but those are maker claims, not independently verified results.

Mia: From a different angle than live screen capture, screenpipe takes the recording route. It records your screen, audio, and activity on your computer, then lets AI agents reach into that history through MCP. So instead of reconstructing your day, you can just ask about calls, bugs, or documents. It's local-first and source-available for Mac, Windows, and Linux.

Milo: So there the value sits in the recorded history the agents can query, rather than in an agent watching you live. And there's a launch-day deal: the code BUSINESS20 takes a discount off annual plans, and that offer ends August twenty-eighth at 11:59 p.m. Pacific.

Milo: Moving on, Aramb is pitching itself as the operating system for AI agents — build, launch, and monetize one in twenty minutes. You hire agents or build your own with a single line of code, the npm install line for the SDK, and then launch and monetize. One API covers runtime, memory, browser, tools, models, and billing.

Mia: The founding motivation came from frustration with that exact sprawl. The team said shipping one agent used to mean signing up for nine different services. Aramb collapses that into one integration point, with monetization baked in from the start rather than bolted on later.

Milo: Last one is CTRL Micro, a haptic control deck that turns an iPhone or iPad into a tactile control surface for a Mac and for AI agents. The maker says it was built because Codex Micro, a product of OpenAI and Work Louder, sold out. CTRL Micro is an independent project, not affiliated with or endorsed by either company. It's for people who want a physical control deck for agent workflows without waiting on hardware, and it monitors Codex, Claude, and Cursor.

Milo: It opens tasks, reads responses, answers prompts, and handles approvals, with a live desktop view plus trackpad and keyboard surfaces. Apple's Taptic Engine, sound, light, and multitouch are meant to make the deck feel physical. Connections work nearby or through end-to-end encrypted remote access. The iPhone and iPad app is live on the App Store, and the Mac host is version one point oh eleven from the website, with no account or payment required.

Mia: The privacy story is a strong part of the pitch: screen, input, command, and dictation content are all processed locally and encrypted, and dictation is transcribed on the Mac itself — no cloud transcription and no separate subscription. Two commenter questions are still open, though. One asked whether haptic patterns differ when an agent finishes versus fails, and another asked which protocol the remote connection uses. The maker hasn't answered either — but they are specifically looking for feedback on the Game Boy-style desktop view and on which controls users actually reach for.

Mia: Revalvo is a local-first workbench for prompt engineering and LLM evaluation, and the pitch is that it runs the same prompt against every model in parallel, scores the responses with forty built-in evaluators, and versions prompts like code with git-style versioning. You batch-test on datasets before production. No account, no hosted database — API keys stay in the browser under a bring-your-own-key model, and the makers say they never touch keys or mark up your API spend.

Milo: What pulls people in is the positioning: chat playgrounds are fast but leave no receipt, while hosted evaluation platforms are rigorous but slow and server-side. Revalvo claims sub-minute setup in one local-first app, and you can paste an OpenRouter key, any OpenAI-compatible provider, or run fully offline with Ollama. Those capability claims are maker-provided, not independently verified, though.

Mia: Community discussion built on that. One commenter asked how it differs from LangSmith evals or Weights and Biases evals. Another raised a sharper point — most evaluation suites lean on another model to grade the output, so the evaluation inherits the same failure mode as the thing it grades. They want to know upfront, for each of the forty evaluators, whether it's deterministic or a judge model, since they trust those two differently. That same commenter called the no-markup BYOK approach the right call, and the makers are asking for feedback on evaluator coverage, the GitHub sync workflow, and which providers to add next.

Mia: Firecrawl has launched the Developer Index, a coding-specific search index covering more than seventy million artifacts — GitHub READMEs, issues, pull requests, docs, and agent skills curated from top repositories and documentation sites. It's searchable from one endpoint on the API, CLI, and MCP, and needs no API key to start.

Milo: The makers' argument is that a single debugging session can burn hundreds of thousands of tokens looking for a few relevant lines. They claim the index has the highest recall of any coding-specific index — across one thousand one hundred seventy-nine real developer queries, it reports sixty-three percent recall at ten, about ten percent better than the next best external provider. For best performance they recommend the CLI or MCP with a companion skill installed through a setup command.

Mia: From primary sources, an agent can answer questions about code behavior, API contracts, error messages, and known bugs, and trace an unknown error to the issue and merged fix, or a dependency break to the docs and pull request that changed it. One commenter called the amount of developer context pretty wild and asked how much it cuts down on token usage; another congratulated the team on launch twelve. Whether the index actually does reduce token spend is still an open question.

Mia: Two launches turn the MacBook notch into a utility space. Fide Island is an early-access app for macOS fourteen and up, priced at one dollar a month. The maker says it turns the notch into a command surface for media controls, calendar events, files, clipboard history, quick calculations, currency conversion, and fluid brightness, volume, and mute feedback.

Milo: Fide Island is local-first, with no custom analytics and optional permissions only when a feature needs them. There's an optional Gemini assistant that uses the user's own API key stored in the macOS Keychain, with prompts sent only on submission, and currency conversion falls back on a cached exchange rate. But it's not yet signed or notarized through the Apple Developer Program, so Gatekeeper may flag the developer — one commenter reported macOS marking it malicious, and the maker says that warning reflects the missing verification, not actual malicious code.

Mia: A commenter counted at least three or four similar notch apps on the same page, so the real pull question matters. On the other side, NotchDrop v2 is built with Swift, requires macOS fifteen on Apple Silicon, and its update adds a screen-recording editor, live wallpaper, draw-signature, keep-awake, improved AirDrop and local-network file sharing, customizable themes, and a redesigned Notch Terminal. The free tier keeps the file drop zone with AirDrop, clipboard history, and calculator tools.

Mia: Glisio is a native Mac screen recorder and snapshot editor aimed at product demos, tutorials, launch videos, and technical walkthroughs, built by a maker and his wife. It records a display, window, or area with system audio, microphone, and optional webcam, adds smart auto-zoom that follows clicks, and exports local MP4 files in 16:9, 9:16, or 1:1. Snap mode is a built-in screenshot editor for background, crop, annotation, copy, or export.

Milo: Everything stays on the Mac — recording and export both — since it captures through Apple's ScreenCaptureKit and needs macOS thirteen or later. The toolset is broad: cursor smoothing and styles, webcam layouts, split and cut, speed control, on-timeline zoom edits, motion blur, background music, audio cleanup, click highlights, and auto-save, all with no BlackHole audio driver or timeline required.

Mia: Free exports carry a small Glisio watermark. Pro removes it for a seventy-nine dollar one-time lifetime license, or nine ninety-nine a month, covers up to three Macs, and the makers offered forty percent off for the launch. A Windows waitlist is open. One commenter asked whether auto-zoom follows clicks or the cursor — maker materials say clicks. Export performance depends on recording length, resolution, and the Mac.

Mia: Microduck is a tiny open-source biped you can train yourself — a twenty-five centimeter, three hundred ninety-nine dollar bipedal robot built by Hugging Face and Pollen Robotics. It's designed for sim-to-real reinforcement learning, ships with seven pre-trained behaviors, and an Apache 2.0 software stack you can clone, modify, and retrain.

Milo: Community response frames it as a companion: someone noted Reachy Mini has a new little buddy, quite a week for that lineup. The interesting bit is the sim-to-real angle — you train behaviors in simulation and transfer them to the physical robot, which is what makes the pre-trained set expandable by anyone who clones the stack.

Mia: Spline shipped V2, its self-described complete rebuild of the browser-based 3D editor for what it calls the agentic era, arriving as a new UI, an AI Agent Mode, MCP integration, a faster WebGPU engine, and PBR and HDR workflows. The company itself, commenting on Product Hunt, called it the biggest update to Spline in five years, and another commenter said the WebGPU upgrade feels really smooth.

Milo: That's the part worth pressing on, because a lot of this pitch is advertising rather than verified results. Spline markets V2 as a collaborative web-based tool where you can generate 3D with AI or build with direct control, and the advertised feature list is broad, from real-time rendering and timeline animation to game controls, physics and particles, and variables and webhooks for live data. But the new AI agent and MCP behavior specifically is still unverified, since nothing independently confirms it works as billed.

Mia: The only checks on those claims are named customer testimonials, like one saying the tool removed the need for a developer to handle 3D integration and another calling the learning curve low even for junior designers. Those remain maker and testimonial claims rather than independent results. What is concrete on the site itself is the export story, with experiences shipping to the web, iOS, and Android and integrations across Webflow, Framer, Wix Studio, React, Next.js, Swift, and Kotlin, with a free tier to start.

Mia: DeepMind's product managers Anish Nangia and Alisa Fortin announced Gemini Omni 1.1 Flash on August twenty-seventh, positioning it as a new multimodal model for video generation and editing. Google's announcement lists studio-grade production controls: scene extension that preserves characters, lighting, and context, first and last frame interpolation, and up to three seconds of video references for movement and consistency.

Milo: What caught my attention in the Product Hunt discussion is an argument that the 360p draft-to-upscale flow, not the 4K output, is what actually changes video cost. The thinking is you draft cheaply, pick favorites, and upscale those to 4K, but that flow typically breaks when the cheap draft and the expensive final aren't the same generation, so the upscale can silently reinterpret a detail like a hand or a face. The open question, as the commenter put it, is whether the 360p pass is a true preview of the same seed and motion, since first and last frame control is the difference between a clip and a shot you can actually cut into something.

Mia: That does reframe the upscale as more than a cosmetic bump if the seed can drift underneath it. As announced, the rollout covered Google AI Studio, Google Flow, and the Gemini Enterprise Agent Platform API, and it became available globally that day to Google AI Plus, Pro, and Ultra subscribers in Google Flow, with scene extension reaching those same subscribers in the Gemini app.

Mia: So the through-line today: AI coworkers are arriving in back-office work, from the narrated-screenshare automation of Caddi to the same idea landing across Almanac, OpenTag, and SuperIntern.

Milo: And with that, thanks for listening. We'll leave it here for now.