ngrok

We're constantly adding support for the latest models and providers at ngrok.ai. We got opus5.5 and gpt6 sol/luna added today. Check it out.

Self-hosted n8n is great until a workflow needs to receive a webhook. A quick tunnel gets you through a demo 👀. But if a Telegram bot, Stripe event, or another external service needs to reach your workflow 24/7, you need a stable public HTTPS endpoint. Ryan Blunden put together a Docker Compose setup that runs n8n and ngrok together, using ngrok’s free static dev domain so the webhook URL survives container restarts. He also walked through the whole setup with a real Telegram webhook. Clone the repo and follow along: https://lnkd.in/gZPf9QyJ Watch the video on YouTube: https://lnkd.in/gJAipVKH

The best compressors are... LLMs? Turns out, "predict the next token" is a pretty useful skill when your job is making text smaller. A compression model assigns probabilities to what comes next. An entropy coder turns those probabilities into bits. High probability, fewer bits. Low probability, more bits. An LLM supplies those same probabilities, with a lot more context to work from. Feed them into an entropy coder and you can compress text losslessly. In Annie Sexton's example, even GPT-2 squeezes a Dickens quote to less than half the size of a simple context model's output. There is, of course, the small matter of running a language model every time someone downloads a stylesheet. Annie walks through the connection in 17 minutes, from shrinking a string into a single number to why gzip still gets to keep its job. Watch the whole wondrous thing on YouTube ↓ https://lnkd.in/gFxpkfMN

GOPHRS is here. AI gateway migration is mandatory. You know, one endpoint to proxy all our requests to all our providers? The arch is final (as in no more feature requests final). Please watch the full architecture briefing before submitting questions. -- Yeah, no. You shouldn’t roll your own. Connect frontier providers and self-hosted models through one managed gateway → https://ngrok.ai Made with the wacky folks at @KRAZAM.

What do a port of Kubernetes to TypeScript and a high-priority bug in the kubelet have in common? They both feature in Sam Rose's guide to probing, out now. We're experimenting with what next-generation k8s education looks like, and we hope you love it. Stick around for the probe playground at the end to dial in the right setup for your own pods ↓ https://lnkd.in/gJ8hTZii

Compression and LLMs are trying to solve the exact same problem: predicting what comes next. Annie Sexton's interactive essay walks us through the basics of compression to reveal its surprising overlap with every language model you've ever used. Check it out → https://lnkd.in/gMSFA-8f

Detailed provider and model usage data is now splashed all over our AI Gateway's dashboard. You'll see: → Requests, success rate, total cost, and average cost per request up top, filterable by provider and time range. → Usage and failure rate charted over time, so you can tell whether an error spike was your code or the provider. → Model, access key, tokens in and out, and cost, down to fractions of a cent, for every attempt. Self-hosted models show the exact same data. A model you're serving with vLLM sits in the provider table right next to OpenAI and Anthropic. Sign up → https://app.ngrok.ai

Every AI gateway can route to OpenAI. The interesting part is routing to the model running on your own hardware. Our own Ryan Blunden demos exactly that: → A translation app calls GPT through the ngrok AI Gateway with a base-URL change and a credential. → He then adds Gemma 4, running in LM Studio on his own machine, as a custom provider. → The local model never gets a public endpoint. The gateway reaches it over a private tunnel. Same endpoint, same SDK, models you rent and models you run. Watch Ryan's demo below, then try it at https://ngrok.ai

Mac Chaffee has argued with coworkers at every job they've had, from their first internship onward. Not long ago, they decided to ask the question out loud: Am I the problem? To find out, they sat down with our infrastructure team to talk about how the team works through disagreement. The conversation landed on a real decision: adopt Argo Rollouts, or keep building rollout tooling on what the team already had. The stakes were not small. Our Mux service holds hundreds of thousands of long-lived TCP connections, and early rollouts ran from a script on a developer's laptop. If the wifi cut out, the deploy failed. Four engineers held four different positions on that decision, and nobody fought about it. Each described the others' viewpoints as accurately as their own. Mac went in expecting conflict and could not find any. The second half of the post is Mac turning the question on themself, with help from the Gottman method, Crucial Conversations, and what mutual aid organizers know about burnout and conflict. Delighted to give Mac this peek into how we work, and, in turn, sponsor their own. Find the answer to Mac's question here: https://lnkd.in/g3G38gVH