Today we're launching Founding Agent. For 25 years, a website's answer to "I have a question" has been a search box and a contact form that goes nowhere. So visitors leave - and you never know they were there. Founding Agent is a voice AI agent that lives on your site and actually talks to people. It answers their real questions instantly, qualifies who's a fit, and books the meeting while they still care. It feels like a real conversation because retrieval runs in under 10ms. No lag, no "thinking" dots. It's built on the same infra 1500+ developers are already using at @usemoss Websites have been one-way mirrors for too long. Time they answered back. Early access 👇

Your voice agent has the context. Why is it still waiting to respond? We’re going after that delay next. Moss Inference. Private beta for teams running voice agents in production. moss.dev/product/inference

Most voice agents fetch context the slow way. The LLM decides it needs something, calls a tool, waits, then answers. That's dead air on every lookup. With the @livekit integration, @usemoss runs retrieval on every user turn before the LLM starts. The context is already in the conversation when it responds. When 80% of our customers build on LiveKit, this wasn't a nice-to-have. It was the obvious next integration.

in la tomorrow. if your voice agent takes longer to reply than you do, let's talk

Something new is coming... 👀 moss.dev/product/inference?v…

co-located

Over the next 6 months, we will see a surplus of personal agents tackling every aspect of personal life.

the thing i respect about notetakers is they are always on time to the meetings

writing emails was one of the simplest things I did pre-ChatGPT and LLMs era... now its the most complex thing

Me when someone books a B2B demo request with a personal email

I came to a point where I automatically ignore any inbound that starts explaining my own startup to me.

Thanks, @srimalireddi ! @usemoss was the missing piece that allowed us to create a truly human-like experience. Thanks to Moss, Kindred answers questions so quickly that it doesn't break the immersion with extra latency or filler words. Our clients love the instant lookup.

When @civiliangame showed me this, neither of us touched a thing. He just talked, and a working voice agent came together exactly the way he described it. Completely hands-free. It felt like magic. It also felt like a preview. For decades, software has asked us to learn it. The menus, the forms, the settings pages. I think that's ending. The next generation of software will just listen. You'll say what you want, and it gets built, booked, or answered while you're still talking. Touch-free won't be a feature. It'll be the default. Proud that @usemoss gets to play a small part under the hood here. Congrats Phillip and the @kindredpm team. Go try it 👇

highly recommend!

OpenAI just priced long context as a luxury good. GPT-6 Astra shipped September 3 with a 1.05M token context window. Past 272,000 input tokens, the bill switches to a second rate card. 2x input and cache rates, 1.5x output. Not on the overflow. On the full request. Here's what that looks like in practice. Send 270,000 input tokens and get 2,000 back, you pay $2.80. Send 280,000 with the same output, you pay $5.75. Four percent more context, twice the bill. Every token you were already paying the cheap rate for gets repriced on the way out. Two weeks ago I wrote that cheap million-token context was the best thing that ever happened to retrieval, because it solved cost and left latency as the real constraint. OpenAI just took the cost part back. And the interesting part is that Astra is genuinely good at long context. 96.3% on MRCR v2 in the 512K to 1M band, up from 73.8% on Sol. The capability is real. The price sheet is telling you not to use it. That's the signal worth reading. A million-token window is not a replacement for retrieval. It's a fallback with a cliff sitting at 26% of the advertised capacity. We built @usemoss on the idea that knowledge should live next to the model, not across a network hop. That used to be a latency argument. It's now in OpenAI's own pricing table. Retrieval isn't a workaround for small context windows. It's the cost floor. The window is capacity. What you choose to send is architecture.

Moss, @usemoss, just crossed 500K+ package installs. Five months ago I posted about hitting 100K. I remember refreshing the PyPI stats page like it was election night. Every install felt like someone saying "I'll give this a shot." At 500K, I don't refresh the stats page anymore. Not because I care less. Because the pattern proved itself. We still haven't run ads. Still no DevRel team. Still no Product Hunt launch. The growth loop hasn't changed since install number one: a developer tries Moss, it works, they tell another developer. What changed is who's installing. At 100K, it was individual developers experimenting. Trying Moss in a sandbox. Seeing if the latency claims were real. At 500K, it's teams. A developer brings it in for one project, their team adopts it, and six months later it's in every pipeline they run. That's the difference between curiosity and trust. The hardest part of building developer infrastructure isn't getting someone to try your product. It's getting them to keep using it after the demo is over. Every install after the first one from the same team is a vote that it works in production, not just in a notebook. 500K installs. Zero paid acquisition. Same growth loop since day one. We're just getting started.

150 countries. 40,000+ production devices. 30 million+ minutes. A year ago none of these numbers existed. We bet early on a product-led motion at @usemoss. No outbound sales team. No enterprise pitch decks. No paid acquisition. Just a product that either worked or didn't, and developers who would tell each other the truth about it. The pattern has been the same from day one: → A developer hits a retrieval bottleneck in their voice pipeline → They find Moss → They test it in a sandbox → They see the latency numbers are real → They deploy to production → Their peer finds out, and the cycle starts again We didn't have a launch moment. We had hundreds of quiet ones. Global Voice AI platforms replacing their knowledge base infrastructure with Moss. A $200M+ ARR company going live with Moss before we even knew about the deployment. A startup increasing its production usage by 2,100% in a single month. The CEO of a unicorn asking our Founding Agent, mid-conversation, what technology was powering it. That's what product-led growth actually looks like. Not a strategy deck. Not a funnel. Just someone with a real problem finding a real solution and deciding to trust it with their production traffic. Every one of those 30 million minutes started with one developer saying "let me try this." Huge gratitude to our early believers. The teams who tried Moss when we were a nimble company with a handful of docs and a big claim. You took a bet on us before anyone else would. To celebrate, we're giving away 1 month free of our Hobbyist tier to anyone who comments "𝗠𝗢𝗦𝗦" on this post. We're just getting started.