Vercel

Vercel is a developer infrastructure company building web application tools, AI-generated software and infrastructure for agents. Its AI products address distinct parts of that work: AI SDK connects applications to models and tools, v0 generates application code, and Sandbox and Workflow SDK provide isolated execution and durable progress. Founded by Guillermo Rauch and formerly branded ZEIT, the company expanded from simplifying deployment into supporting the development and operation of applications and agents.[2][1][9][8][4]

At Ship 2026, Vercel organized its strategy around three roles: infrastructure on which coding agents deploy software, tools for building and running agents, and agents that help operate Vercel itself. Rauch remained CEO and Malte Ubl CTO. The event presented eve, an open-source agent framework, and announced AI SDK 7.[11]

The products divide the work into concrete responsibilities. AI SDK supplies model calls, streaming, structured output and tool loops through a provider-agnostic TypeScript interface. Developers can connect directly to providers or use AI Gateway, which handles authentication, routing, usage tracking, failover and billing. Sandbox runs generated code in isolated Linux microVMs. Chat SDK connects agents to communication platforms, while Vercel Connect supplies scoped, short-lived credentials for third-party services.[9][8]

Workflow SDK addresses work that must outlive its initiating request. In Vercel’s agent guide, a flight-search example uses `WorkflowAgent` to turn tool calls into persisted steps. Failed steps retry from checkpoints, and tools marked `needsApproval` can suspend execution until a person responds. The guide recommends starting with an in-memory agent and adding durability when work or approvals outlast request timeouts.[8]

In a July 2026 interview, Chief of Software Andrew Qu traced eve to recurring problems in v0 and internal agents: “switching models or providers, adding fallbacks and making runs resumable.” He said reusable solutions accumulated into a more cohesive framework. In eve, agents occupy a directory containing Markdown instructions and TypeScript tools, with durable execution, sandboxed compute, approvals, subagents and evaluations already integrated.[10][11]

Vercel’s agent infrastructure followed an earlier investment in web frameworks and deployment workflows. That history began under the ZEIT name, before AI SDK and v0.[1][3][4]

In his April 2020 rebranding announcement, Rauch dated the company’s beginnings to 2016 and described its aim as “empowering solo developers to effortlessly deploy their apps.” Its first product, `now`, made applications and services available through a single command, including SSL and built-in routing. It simplified the work of getting software online across domains, networking and deployment infrastructure.[1]

ZEIT became Vercel on April 21, 2020, alongside a stronger emphasis on Develop, Preview, Ship: a team’s cycle of implementation, feedback and release. Next.js, its open-source React framework, was already a central product. Vercel also announced $21 million in Series A funding, led by Accel and CRV, to support the platform and continued open-source investment. Existing workflows, pricing plans and projects remained unchanged at the rebrand.[1]

At the rebrand, Vercel reported that Next.js powered more than 35,000 sites, including sites at Uber, Nike and Starbucks, and that more than 300,000 frontend engineers and designers used its technology. These were company-reported adoption figures in April 2020, rather than counts of paying customers.[1]

Vercel’s AI SDK launched on June 15, 2023, with an announcement credited to Jared Palmer, Shu Ding and Max Leiter. The open-source JavaScript and TypeScript library initially supported React/Next.js and Svelte/SvelteKit, with adapters for OpenAI, LangChain and Hugging Face Inference. Its `useChat` and `useCompletion` hooks helped applications render text as it arrived; callbacks supported saving prompts, tokens and completed responses. The toolkit later expanded to model calls and tool-running agents through interfaces such as `generateText` and `ToolLoopAgent`.[1][3][9]ⁿ¹

v0 generated the interface itself. By its October 11, 2023 beta transition, users could describe a UI, receive React code using Tailwind CSS and shadcn/ui, select an iteration, edit it and copy the code into an application. Palmer described its purpose as helping developers build a product’s first iteration. Vercel reported 100,000 waitlist registrations in three weeks and was extending access to 5,000 additional users; registrations did not establish active usage.[4]

Vercel’s June 2025 composite-model architecture surrounded v0’s base model with specialized components. Preprocessing included recent conversation messages, summaries of older messages, and retrieved documentation, UI examples and project sources. Large changes went to a frontier model; narrow edits went to Quick Edit. AutoFix checked output during streaming, followed by a final correction and lint pass. Separating these tasks allowed Vercel to replace the base model without rebuilding the pipeline.[5]

Early correction pipelines used Gemini Flash 2.0. Vercel subsequently trained `vercel-autofixer-01` with Fireworks AI, using reinforcement fine-tuning across tracked error categories. On its internal web-development evaluation, Vercel reported an error-free generation rate of 93.87% for v0-1.5-md, compared with 64.71% for Claude 4 Sonnet. These results measured the company’s defined tasks and errors, rather than general software correctness; the architecture and model selections describe the published 2025 system.[5]

d0 is Vercel’s internal Slack agent for answering data questions with SQL. In December 2025, Qu described a redesign after specialized retrieval, validation and planning tools became a maintenance burden. The new version gave Claude Opus 4.5 access through Sandbox to files documenting the Cube semantic layer—dimension definitions, measure calculations and join relationships. It explored YAML, Markdown and JSON with standard Unix tools, while retaining a separate SQL execution tool.[6]

Across five company-selected representative queries, Vercel reported five successful answers from the redesign versus four from its predecessor, with mean execution time falling from 274.8 to 77.4 seconds. The result was a small internal comparison. Qu attached a specific condition to the approach: “This only worked because our semantic layer was already good documentation.” Clear definitions and relationships let the model inspect domain knowledge directly instead of relying on tools that summarized it.[6]

In his February 2026 account, Qu recalled Shu Ding writing down React and browser expertise over a weekend the previous year. They considered a blog post or documentation, and an MCP server, before choosing skills to deliver the Markdown knowledge on demand. Repeating installation instructions for different coding agents led Qu to build `npx skills`, a shared installer. Installation telemetry then supplied the leaderboard at skills.sh, which made skills discoverable as well as installable.[7]

By February 20, 2026, Vercel reported more than 69,000 skills and 2 million skills CLI installs. As the ecosystem grew, it announced audit partnerships with Gen, Socket and Snyk, along with plans for per-skill assessments. The counts described skills and installer activity, rather than unique developers or Vercel customers.[7]

Vercel’s June 2026 announcements also brought agents into platform operations. Vercel Agent, then in public beta, investigated production alerts and anomalies and proposed fixes through pull requests. It ran under its own identity, was read-only by default and requested narrow, temporary permissions for production actions. Qu’s July interview favored feedback cycles fitted to the task: well-defined work could run autonomously, while careful engineering needed human steering. He identified shared context between teammates as a further problem to solve.[11][10]