Five products launched yesterday, and one of them puts your logo on boxer shorts

September 11 brought a strange but useful spread of new developer tools. A gateway that swallows 31 model providers, an IDE that turns your terminals into a map you can zoom, Meta’s take on a personal agent, a browser-based subtitle studio, and an ad network with exactly one inventory slot. No ranking data had closed when I pulled this together, so treat the order below as my own read rather than a community verdict.

SubtitleGenerator

Subtitles are tedious in a specific way: transcription tools get you 90 percent there, then you spend an hour hunting the words they got wrong. SubtitleGenerator runs entirely in the browser with Next.js, flags uncertain words for your review, and ships all 33 caption styles on the free tier. Pro and Max add full-track translation, watermark-free HD exports, more formats, custom fonts and saved brand styles.

The detail worth noting is that your video never leaves the browser, which matters if you’re working with client footage or anything under embargo. The honest caveat is that the free tier is styling and correction, not translation or clean HD export. If you only need an SRT file for YouTube, you may never hit the paywall. Podcast clippers and solo creators dealing with messy audio will get the most out of it.

Ass Auction

The premise is genuinely narrow: one leaderboard, real money, no accounts, and the top 22 bidders get their logos worn on a pair of boxers. A minimum of $5 buys a position, Stripe handles payment, and rank is simply total spend. The board publishes a steal price against each entry, so you can see exactly what it costs to overtake the person above you. Money is never refunded, and the site says so without apology.

It’s not for everyone. Click counts captured at launch were modest, with six claimed entries sending a combined 158 clicks, so expect inquisitive traffic rather than a campaign-scale audience. Brand-safety teams will hate it. Indie makers with a sense of humor and $5 to spare get a permanent public link, a share card, and an email whenever someone knocks them down the list. The gossip bar running alongside is a nice touch that turns the audience into part of the show.

Muse by Meta

Meta’s agent is the most ambitious thing on this list and the hardest to judge without using it. Muse runs on a persistent, isolated Linux VM with a full browser and real CPU, memory and storage, so it can keep working across sessions rather than resetting each time. You talk to it through the Muse app or WhatsApp, which lowers the adoption barrier for anyone who already lives in a chat window.

The control layer is what makes it interesting. Actions like sending emails and making purchases require your approval, everything gets written to an audit trail, logins sit in a credential store the agent cannot read, and shopping checkouts generate one-time card numbers. Conversations aren’t shared with Meta’s ad systems. The listed price is free, though the site doesn’t explain limits, and the scope on offer, from finances and health to shopping and relationships, is broad enough that I’d want to see it working before trusting it with anything that touches money.

49agents IDE

Tab fatigue is real when you run several coding agents at once. 49Agents replaces the stacked terminal tab with an infinite 2D canvas where every terminal, repository, agent and machine occupies a stable position. Panes can be terminals, notes, files, git graphs, web pages or directories, and they glow blue while Claude is working and pulse vermillion when it needs permission. Remote machines live on the same surface with live CPU and RAM readouts.

Setup is three commands, and it opens on localhost:1071 with no account or token. You need Node 18 or higher, tmux and ttyd. The catch is obvious: if you run one terminal, you don’t need this. For anyone juggling half a dozen agent sessions across a laptop, a home desktop and a cloud box, remembering where things sit beats reading tab titles.

GoModel

An open-source AI gateway in Go that puts a single OpenAI- and Anthropic-compatible endpoint in front of 31 providers. Point your SDK at it, change the base URL, and you get routing, failover, budgets, guardrails, exact and semantic caching, audit logs and per-request cost attribution. It ships as one binary with an embedded dashboard, under MIT, with storage starting on SQLite and moving to Postgres or Mongo as traffic grows.

The benchmark claims are striking: 2.35 ms median overhead against 42.4 ms, and 3,610 requests per second against 250. Realistically, it’s heavier to adopt than the site implies. Semantic caching wants a vector database, and the scoped workflow system takes time to learn. Platform teams tired of per-provider integration code will still find it the most complete release of the day.

Yesterday's Top Launches: 5 Tools from September 11, 2026