AI agents are improving at handling complex tasks, but a recurring limitation emerges when they need to access paid external services. The current model requires pre-registering with every API, managing keys, and maintaining separate billing relationships. This workflow breaks down if an agent needs to autonomously discover and interact with new services. x402 is an open protocol that addresses this by adding payment directly into HTTP requests. It uses the 402 Payment Required status code to let any API endpoint request payment without prior account setup. We built x402-mcp to integrate x402 payments with Model Context Protocol (MCP) servers and the Vercel AI SDK.

MongoDB Atlas is now available on the Vercel Marketplace. Developers can now provision a fully managed MongoDB database directly from your Vercel dashboard and connect it to your project without leaving the platform. Adding a database to your project typically means managing another account, working through connection setup, and coordinating billing across services. The Vercel Marketplace brings these tools into your existing workflow, so you can focus on building rather than configuring.

We’re updating Vercel’s Pro plan to better align with how modern teams collaborate, how applications consume infrastructure, and how workloads are evolving with AI. Concretely, we’re making the following changes:

When the MCP standard first launched, many teams rushed to ship something. Many servers ended up as thin wrappers around existing APIs with minimal changes. A quick way to say "we support MCP". At the time, this made sense. MCP was new, teams wanted to get something out quickly, and the obvious approach was mirroring existing API structures. Why reinvent when you could repackage? But the problem with this approach is LLMs don’t work like developers. They don’t reuse past code or keep long term state. Each conversation starts fresh. LLMs have to rediscover which tools exist, how to use them, and in what order. With low level API wrappers, this leads to repeated orchestration, inconsistent behavior, and wasted effort as LLMs repeatedly solve the same puzzles. MCP works best when tools handle complete user intentions rather than exposing individual API operations. One tool that deploys a project end-to-end works better than four tools that each handle a piece of the deployment pipeline.

On September 9, 2025, the campaign extended to DuckDB-related packages after the duckdb_admin account was breached. These releases contained the same wallet-drainer malware, confirming this was part of a coordinated effort targeting prominent npm maintainers. While Vercel customers were not impacted by the DuckDB incident, we continue to track activity across the npm ecosystem with our partners to ensure deployments on Vercel remain secure by default.

Recently we partnered with the Biome team to strengthen their noFloatingPromises lint rule to catch more subtle edge cases. This rule prevents unhandled Promises, which can cause silent errors and unpredictable behavior. Once Biome had an early version ready, they asked if we could help stress test it with some test cases. At Vercel, we know good tests require creativity just as much as attention to detail. To ensure strong coverage, we wanted to stretch the rule to its limits and so we thought it would be fun to turn this into a friendly internal competition. Who could come up with the trickiest examples that would still break the updated lint rule? Part of the fun was learning together, but before we dive into the snippets, let’s revisit what makes a Promise “float”.

At Vercel, our relationship with open source is foundational. We do not build open source software to make money. Rather, we’re building an enduring business that enables us to continue developing great open source software. We believe in improving the default quality of software for everyone, everywhere, whether they are Vercel customers or not. A rising tide lifts all boats.

Many engineering teams have disaster recovery plans. But unless those plans are regularly exercised on production workloads, they don’t mean much. Real resilience comes from verifying that systems remain stable under pressure. Not just in theory, but in practice. On July 24, 2025, we successfully performed a full production failover of our core control-plane database from Azure West US to East US 2 with zero customer impact. This was a test across all control-plane traffic: every API request, every background job, every deployment and build operation. Preview and development traffic routing was affected, though our production CDN traffic, served by a separate globally-replicated DynamoDB architecture, remained completely isolated and unaffected across our 19 regions. This operation was a deliberate, high-stakes exercise. We wanted to ensure that if the primary region became unavailable, our systems could continue functioning with minimal disruption. The result: a successful failover with zero customer downtime, no degraded performance in production, and no postmortem needed.

Prototyping with AI should feel fast, collaborative, and on brand. Most AI tools have cracked the "fast" and "collaborative" parts, but can struggle with feeling "on-brand". This disconnect usually stems from a lack of context. For v0 to produce output that looks and feels right, it needs to understand your components. That includes how things should look, how they should behave, how they work together, and all of the other nuances. Most design systems aren’t built to support that kind of reasoning. However, a design system built for AI enables you to generate brand-aware prototypes that look and feel production ready. Let's look at why giving v0 this context creates on-brand prototypes and how you can get started.

Building an AI app can now take just minutes. With developer tools like the AI SDK, teams can build both AI frontends and backends that accept prompts and context, reason with an LLM, call actions, and stream back results. But going to production requires reliability and stability at scale. Teams that connect directly to a single LLM provider for inference create a fragile dependency: if that provider goes down or hits rate limits, so does the app. As AI workloads become mission-critical, the focus shifts from integration to reliability and consistent model access. Fortunately, there's a better way to run. AI Gateway, now generally available, ensures availability when a provider fails, avoiding low rate limits and providing consistent reliability for AI workloads. It's the same system that has powered v0.app for millions of users, now battle-tested, stable, and ready for production for our customers.

Code and Theory is a digital-first creative and technology agency that blends strategy, design, and engineering. With a team structure split evenly between creatives and engineers, the agency builds systems for global brands like Microsoft, Amazon, and NBC that span media, ecommerce, and enterprise tooling. With their focus on delivering expressive, scalable digital experiences, the team uses v0 to shorten the path from idea to working software.

How do you tell an AI agent what it needs to do when it hits a protected page? Most systems rely on external documentation or pre-configured knowledge, but there's a simpler approach. What if the instructions were right there in the HTML response? llms.txt is an emerging standard for making content such as docs available for direct consumption by AIs. We’re proposing a convention to include such content directly in HTML responses as <script type="text/llms.txt">.