June 2025 Summaries

Vercel Ship 2025 highlighted the company's advancements in app development with a focus on AI, compute, and security, as over 1,200 participants gathered in New York City for the event. Key innovations included the AI Gateway, which streamlines access to multiple AI models without vendor lock-in, and Fluid compute, which optimizes serverless workloads to significantly reduce costs. The introduction of Vercel Sandbox provides a secure environment for running untrusted code, while Rolling Releases ensure safer deployments with real-time monitoring and incremental rollouts. Vercel also showcased Microfrontends for team autonomy, BotID for invisible bot detection, and the Vercel Agent for enhanced observability. These advancements aim to transition Vercel from a Frontend Cloud to an AI Cloud, supporting both developers and AI agents in building and optimizing AI-powered applications.

Bots have become increasingly sophisticated, able to mimic human-like behaviors such as executing JavaScript and solving CAPTCHAs, making traditional defenses like header checks ineffective. BotID, a new layer of protection available on Vercel, acts as an invisible CAPTCHA to safeguard critical routes from automated abuse, such as checkouts, logins, and APIs. It offers two modes: Basic, which is enabled by default, and Deep Analysis, which provides advanced detection capabilities powered by Kasada. BotID operates by injecting obfuscated code that evolves with each load, collecting thousands of signals to differentiate human users from bots without altering user experience. This system requires minimal setup and no configuration, relying on a global machine learning network to enhance protection continuously. Designed for teams of all sizes, BotID aims to counteract advanced bots that resemble real users, offering a sophisticated layer of defense against automation targeting sensitive parts of applications.

Fluid compute, a new model for handling I/O bound workloads like AI inference and MCP servers, has emerged as an efficient solution for modern application needs, particularly on the Vercel platform. It deviates from traditional serverless models by orchestrating multiple concurrent requests on shared resources, eliminating cold starts and reusing idle time, leading to substantial cost savings and improved efficiency. This approach has become the default for AI workloads on Vercel, resulting in cost reductions of up to 90% due to intelligent resource sharing. To further enhance cost-efficiency, Vercel has introduced an Active CPU pricing model, which charges for CPU usage only when code is actively executing, aligning costs with actual compute work rather than wall time. This model, alongside usage-based billing for provisioned memory and invocations, significantly reduces costs, particularly for workloads with high idle time. The Fluid platform is designed to support modern applications with standard runtimes like Node.js and Python, optimizing infrastructure while maintaining flexibility and performance. The Active CPU pricing model is now available by default for most Vercel users, promising to further optimize costs and resource usage for AI-powered applications.

Modern bots are increasingly sophisticated, mimicking human behavior by executing JavaScript, solving CAPTCHAs, and navigating interfaces like real users, which makes traditional defenses inadequate. BotID is introduced as a new protective layer on Vercel, acting like an invisible CAPTCHA to prevent browser automation from reaching critical backend operations such as checkouts, logins, and API interactions. It operates in two modes, Basic and Deep Analysis, and employs advanced detection techniques without relying on static signals like user-agent headers or IP addresses, thereby reducing the risk of false positives and enhancing user experience. BotID utilizes a dynamic detection logic that evolves with each load, making it resistant to reverse engineering, and integrates seamlessly into the Vercel platform, providing robust protection against automated threats. This solution is designed to protect critical application routes from sophisticated bots that imitate real users, offering fast, reliable, and developer-friendly security without requiring extensive setup or configuration.

Vercel has expanded its partnership with WPP, integrating AI tools such as v0 and AI SDK into WPP's global creative teams, enabling faster design, prototyping, and deployment of digital experiences. This collaboration allows creative teams to reduce production times significantly, with WPP's internal pilots showing a 25% increase in digital production speeds. The partnership, which began in early 2024, has helped agency clients modernize their experiences, allowing for faster and more efficient content updates and feature deployments, while dissolving traditional role boundaries, enabling designers to blend UX, UI, and coding seamlessly. Agencies like VML North America, AKQA, and Ogilvy have benefited from adopting Next.js and Vercel, achieving notable improvements in site performance, cost reduction, and increased efficiency. Additionally, the partnership facilitates community-building and innovation, with initiatives like the WPP x Vercel AI Generative CX Hackathon, demonstrating the transformative potential of AI in digital experiences.

Vercel is expanding its capabilities to embrace the potential of AI while maintaining its foundational principles, transitioning from a frontend focus to a full-stack approach, and enhancing its enterprise offerings. The company has appointed Keith Messick as its first Chief Marketing Officer to support this growth and strengthen the developer community's voice. Keith, with a rich background in marketing to developers, previously served as CMO at Redis and played significant roles at LaunchDarkly, DialPad, and Lucidworks. He combines commercial insight, creative vision, and a technical mindset, and will lead Vercel's marketing team, known for its innovative product launches like AI Gateway, Vercel Blob GA, and Chat SDK. As Vercel continues to evolve, it seeks to deepen its role in the intersection of creativity and technical innovation, with open roles for various marketing positions to further shape its story in AI and web experiences.

The Tray.ai company migrated their marketing site to Vercel with the help of Roboto Studio, aiming to simplify infrastructure, improve developer experience, and consolidate thousands of pages. The old architecture was held together by legacy decisions, including an older version of Next.js and a rigid dev–stage–prod pipeline that slowed development. After the migration, Tray.ai achieved faster rollouts, more agile updates, and reduced build times from a full day to just two minutes. They implemented a hybrid approach with Static Site Generation, Server-Side Rendering, and Client-Side Rendering, and rebuilt their forms engine using Vercel Functions. The new architecture provides real-time visibility into changes, improved reliability, and resilience at the edge, allowing Tray.ai to deploy faster, validate ideas quicker, and debug issues more efficiently.

The Model Context Protocol (MCP) standardizes how to build integrations for AI models, and a new adapter has been created to help developers create their own MCP servers using popular frameworks such as Next.js, Nuxt, and SvelteKit. The adoption of MCP has grown significantly, with companies like Zapier, Composio, Vapi, and Solana deploying their own MCP servers on Vercel, resulting in substantial growth. However, scaling MCP servers reveals limitations in the original design, particularly with the inefficient Stream-Sent Events (SSE) transport method, which has been replaced by the more efficient Streamable HTTP protocol in a new version of the MCP specification. The MCP adapter provides built-in support for both Streamable HTTP and SSE, allowing server developers to choose the most efficient option. With the introduction of Streamable HTTP, MCP servers can now be optimized for efficiency, and forward compatibility is ensured through the use of proxy packages like mcp-remote. The MCP ecosystem continues to evolve quickly, with Fluid compute and the MCP adapter enabling developers to build MCP servers that support both current and future clients.

The conference platform and visuals for Vercel Ship 2025 were designed using ferrofluid-inspired 3D visuals and generative AI workflows. The design team used tools like Flux, Google Veo 2, Runway, and Ideogram to generate over 15,000 images and videos during the exploration phase. A modular component system was implemented to simplify the stack, allowing teams to move independently. Incremental Static Regeneration (ISR) and Partial Prerendering (PPR) were used to improve speed and performance. The Payload admin panel lives on the same domain as the Vercel platform, making debugging and logging more streamlined. A library of reusable, themeable components was built using shadcn, Geist, and cva to support multiple events each year. The site's design was influenced by the need for a fast, fluid, and collaborative experience, with a focus on simplifying the stack and improving rendering strategies and caching.

The text discusses the changing landscape of search engines, particularly with the rise of AI-first interfaces like ChatGPT and Google's AI Overviews. Large language models (LLMs) are now a key layer in the discovery process, favoring content that explains things clearly, deeply, and with structure. This shift requires an adaptation of traditional search engine optimization (SEO) strategies, as LLMs tend to ignore keyword stuffing and hidden text. Instead, creating content that is structured, original, and relevant is crucial for both human searchers and AI systems. The article outlines principles and practices for balancing traditional SEO and LLM SEO, including finding a frontier concept, publishing the definitive source, structuring for machines, seeding authentic citations, and setting a refresh cadence. It also highlights the importance of tracking AI impact and adapting to an evolving landscape where no single metric confirms usage.

An attacker can exploit a language model by injecting malicious commands into its system prompt or tools, which can lead to various security risks such as data leaks and unauthorized behavior. To mitigate these risks, it's essential to design for worst-case scenarios, assume total compromise of the attack surface, and scope tools tightly to prevent unauthorized access. Moreover, proper authorization, validation, and sanitization are crucial to prevent prompt injection and exfiltration through model output. The key is to limit the consequences of an incorrect behavior by treating model output as untrusted by default and avoiding rendering markdown or HTML directly. By designing for failure first and shipping later, developers can minimize damage when their language models behave incorrectly.

AI agents, which are language models equipped with system prompts and tools, face significant security risks, primarily from prompt injection attacks. These attacks are akin to SQL injection, allowing malicious inputs to be embedded within seemingly normal data, thereby compromising the agent's operations. Developers are advised to assume total compromise of the prompt, ensuring tools are scoped strictly to user authority and designing systems with the possibility of every input being compromised in mind. Prompt injections can stem from indirect inputs like database content or web-scraped data, posing a threat even when tools are properly authorized. Exfiltration risks also exist through model outputs, such as rendering injected markdown that can trigger unintended data leaks. To mitigate these risks, developers should treat model outputs as untrusted by default, sanitize outputs before rendering, and apply additional security measures like CSP rules and specialized packages for secure markdown handling. Building agents should focus on minimizing the impact of failures rather than trusting the model to adhere to rules, emphasizing the importance of designing for security from the outset.

Building an effective AI agent requires a nuanced approach, starting with manually simulating the task using real inputs to understand its complexity and identify areas for automation. Once a viable simulation is established, building code that automates the loop of collecting inputs, performing deterministic computation, calling the model when necessary, evaluating results, and deciding on next steps can be done. The key to success lies in focusing on reliability through refinement of prompts, precise tool calls, and iterative testing, ensuring quality holds up across a range of real-world inputs and edge cases. By leveraging regular programming fundamentals with judicious use of LLMs for parts that require judgment, developers can create powerful AI agents that feel like magic but are grounded in smart systems with reasoning built in.

The v0 model architecture combines specialized knowledge from retrieval-augmented generation (RAG), reasoning from state-of-the-art large language models (LLMs), and error fixing from a custom streaming post-processing model to achieve significantly higher quality in code generation. This composite model allows for quick upgrades to the latest frontier model while keeping the rest of the architecture stable, enabling v0 to tackle complex tasks like building full-stack web applications with ease. The model's pre-processing step retrieves additional context based on user queries from a dataset, and its state-of-the-art base models handle generation tasks, with smaller edits routed to an optimized Quick Edit model for speed. A custom AutoFix model is used to catch and fix errors mid-stream, improving output quality. v0 models substantially outperform their base model counterparts in error-free code generation, and the development team plans to continue improving model output and releasing new model classes in the upcoming months.