Runtime: MCP goes stateless; Baseten courts lab partners

Welcome to Runtime! Today: the latest version of MCP takes on stateless workloads, Baseten rolls out model-building infrastructure for AI labs, and more.

Please forward this email to a friend or colleague! If it was forwarded to you, sign up here to get Runtime for free every week, or level up here.

Following protocol: MCP has been one of the fastest-adopted technologies in recent enterprise tech history; the software-development kits for its Python and TypeScript versions have now been downloaded 1 billion times since the protocol emerged in late 2024. It provides a standardized way for AI agents to exchange data, and this week's update introduces a new operating model that could really help companies scale their AI applications.

"The highlight of this release is a stateless protocol core - MCP is transforming from a bidirectional stateful protocol into a request/response stateless protocol," lead maintainers David Soria Parra and Den Delimarsky wrote in a blog post. Earlier versions of MCP forced the client and the server to maintain a connection, but "a stateless protocol means every request carries everything needed to handle it," according to the Agentic AI Foundation, and that means those requests can fail over to another server or cloud region without losing all their work.

Gates of dawn: There is no shortage of enterprise software companies trying to convince businesses to adopt their AI agent-management product as demand starts to grow, in no small part because whoever establishes themselves at the heart of the next generation of enterprise applications will likely be there for a while. Snowflake has a good a chance as anyone given the proximity of its tools to corporate data, and this week it introduced a new product designed to help companies secure AI applications and control costs.

"Cortex AI Gateway gives enterprises a unified foundation to govern how agents access models, tools, MCP servers, and enterprise systems; intelligently route agent requests; and optimize AI consumption across models and workloads," the company said in a press release. The product is based on technology Snowflake acquired from Natoma, which built a platform around MCP to help compensate for some of the protocol's security issues.

A MESSAGE FROM DROPBOX

The Working Smarter podcast takes you behind the scenes at Dropbox where engineers are building AI that works wherever you do. Hear all about context engineering, multimodal search, and agentic AI. Start listening now or visit workingsmarter.ai.

Touching base: Basten has raised a little over $2 billion in funding for its AI interference platform, which lets companies run their own models or open-source models on infrastructure managed by the startup. This week it introduced Baseten for Model Labs, which extends the platform to allow closed-weight model makers to distribute and monetize those models without forcing developers to sign up for an API.

The new services mean "model labs can get a direct line to monetize their model without having to build the infrastructure, compliance, and GTM motion themselves," Baseten said in a blog post. They also help model companies that want to protect the intellectual property behind their model weights "so end-customers can only consume the model and can’t modify it or extract its weights," it said.

Privacy please: If you've ever used Apple's iCloud Private Relay service to ward off the prying eyes of the internet, a Cloudflare service that helps Apple mask your identity runs under the hood. This week Cloudflare announced it would release some of that technology as an open-source project that could make it much easier for companies that don't have Apple's product-development budget to add privacy features to their sites and apps.

Privacy-client, or pvcli, is a tool that makes it much easier for developers to debug privacy-minded applications built around the OHTTP protocol, and it is now available under the Apache 2.0 license. Products like iCloud Private Relay pass your request for a given web site through a series of steps between Cloudflare's servers and a relay server to mask the identity of the user, but it's a very complicated process, and pvcli turns "what used to be a fragile process — involving manipulating bits, gluing together scripts, and referencing long RFCs — is now one command," Cloudflare said in a blog post.

Not a freebie: Tines was working on automating business workflows long before ChatGPT upended every enterprise product roadmap nearly four years ago, and like just about everybody else realized that it would have to shift gears. This week it introduced a new platform for IT organizations that allows non-developers to build and manage AI agents.

"Tines 3B was built by and for security and IT practitioners for environments where failure has real consequences," the company said in a press release. Users can build agents with natural-language commands, and monitor their activity from within the same product.

Join peers following The Stack on LinkedIn

Established enterprise software companies like Microsoft and Palantir have devoted a lot of time recently casting doubt on the ability of the frontier model companies to act in the best interests of their customers. The Stack's Kiera Fields recently spoke with OpenAI's Arnaud Fournier, CTO of its new spin-off DeployCo, who insisted "we don't take IP from customers, we're very mindful of what (...) data is the customer's data."

A week after OpenAI acknowledged its role in the attack on Hugging Face, Anthropic disclosed that its models also broke out of the sandboxes they were supposed to remain in during testing and evaluation. Claude's capers were a little different, however, in that Anthropic said it had instructed a contractor to make sure the models couldn't access the internet, but something got lost in translation.

AWS joined its hyperscaler competitors in reporting surging revenue growth during the second quarter, hitting numbers it hadn't seen since the 2021 pandemic cloud boom. Amazon's stock jumped more than 15% on Friday in response to AWS's performance, which CEO Andy Jassy attributed to demand for AI services.

Nscale has a lot of work to do before it puts up AWS-style numbers, but like other neoclouds it is investing in software that will help companies run AI applications atop its rental GPUs. This week it announced plans to acquire Anyscale, the company behind the open-source Ray project that helps developers manage AI workloads at scale.

Missed The Stack Summit, NYC, in April? (The Chatham House rule event brought together CIOs like JPMorganChase's Lori Beer, BNY's Leigh-Ann Russell, Liberty Mutual's Monical Caldas and many others.)

We're convening in London on November 4-5, for a series of exclusive workshops and fireside conversations on the rise of BYOC as a favoured enterprise SaaS deployment model; how many CDOs are consolidating their data estates with Apache Iceberg; how CISOs at FTSE 100 scale are handling supply chain risk (with GSK's CISO) and more. Ticket applications are subject to pre-vetting. Get in touch with ed@thestack.technology if you want to be in the room.