TOP LAUNCHES

Five AI tools you may have missed

Browse.sh — Browserbase's new product gives AI agents persistent memory for web automation: rather than scripting each interaction from scratch, agents build and replay learned patterns. It was the #2 launch on Monday with 705 votes. The pain of stateless browser automation is apparently widely felt.

Minimi — An ambient memory layer for Claude from the team at Shram: it runs alongside your conversations and retains context that would otherwise disappear at the end of a session. Context windows are getting longer, but memory that persists across sessions is a different problem.

Vaani — Lip-synced AI dubbing in 40+ languages with voice cloning and music preservation. The technical bar is harder than it looks: frame-accurate lip sync while preserving the original voice is the part that separates this from generic tools. Built for creators and studios distributing content globally.

ZeroGPU — An inference routing layer that sends 70–80% of production tasks to smaller edge-optimized models instead of frontier APIs, at 10x the speed and half the cost. The bet: most inference tasks are overprovisioned for the model they're hitting.

Manus Shopify Connector — The AI agent platform Manus launched a vertical-specific product this week: build and manage Shopify stores from a single chat interface. General AI agents are going vertical, and commerce is an obvious first destination.

SPONSORED BY

We asked 34 customers what Viktor does for them. Not one said chatbot.

They kept using words like colleague, coworker, team member. One CEO called it the glue holding their e-commerce business together, which is a lot, but also… you see why.

It lives in Slack and plugs into 3,000+ tools, so instead of jumping between tabs, you just ask for the thing. Pull Stripe against HubSpot, check Sentry alerts, spin up a campaign brief, build a landing page, send a report upstairs. It all happens there.

It has already hit top 5 on Product Hunt with 130 comments, is SOC 2 certified, and your data does not train models.

One user said it was the first time AI felt like a real coworker, which is either exciting or slightly concerning depending on your week.

WHAT’S HOT

Here come the invoices

GitHub Copilot switched to usage-based billing on June 1st. If you haven't noticed yet, you will when the invoice arrives. Copilot was the product that trained most developers to treat AI assistance as always-on infrastructure: always running, never something you'd stop and calculate. That era is done.

The response has been pretty clear. Developers are either rerouting tasks to cheaper models, leaving for Cursor, or quietly accepting the higher bill. The third group is smaller than you'd expect. Nobody voted for hard spend caps when the question was polled in the forum this week, which is interesting. Developers don't want to feel restricted. They just want the math to work.

What's actually happening is that "unlimited" was always a fiction. GitHub was absorbing those compute costs whether you saw them or not. The pricing change doesn't create a new cost. It moves it from GitHub's balance sheet to yours, and suddenly every autocomplete has a number attached. That changes how you use the tool.

The thing it'll probably teach people: you don't need the same model for everything. A boilerplate function is not a hard problem. A regex isn't either. Most of what Copilot does all day doesn't need a frontier model, and developers who figure that out now will spend a lot less than the ones who keep treating every task the same. Copilot accidentally made this an interesting question to answer.