Vercel Eve: adding an AI agent to your Next.js website

Vercel Eve, the open-source AI agent framework explained

Eve is an open-source framework that Vercel released in June 2026 for building AI agents meant to run in production rather than only in demos. Its central idea is that an agent is a folder of files, with a Markdown file for the instructions, TypeScript files for the tools the agent can call, and further Markdown files for skills loaded on demand. Eve compiles that structure and handles everything around it automatically, with no manual configuration to write.

What eve does well

An AI agent that survives failures

Each session runs as a durable workflow, so if the process crashes or the user takes a long time to reply, the session resumes exactly where it stopped, with its context intact.

An AI agent whose code stays isolated

Code that the agent generates itself, such as scripts or shell commands, runs in a sandboxed environment kept apart from your application's runtime. This is a necessary precaution, since code produced by a model should never be treated as trusted code.

An AI agent that waits for human approval

Any action can be configured to require approval before it runs, which prevents an agent from deleting data or sending a message without supervision.

An AI agent that is fully traceable

Every run produces a detailed trace, model call by model call and command by command, which can be exported in the standard OpenTelemetry format to tools such as Datadog or Honeycomb.

Open source, though tied to Vercel by default

This point calls for precision, because the messaging around eve puts a lot of weight on its open-source status. That status is accurate as far as the licence goes, and less so as far as the actual developer experience goes.

By default, eve relies on several proprietary Vercel services, namely Vercel Sandbox to isolate generated code, Vercel Workflows for durable execution and Vercel Connect to manage OAuth connections to third-party services. Vercel's CTO has himself acknowledged that the adapters needed to run on other infrastructure, AWS or Cloudflare in particular, are still in development, with availability targeted for late 2026.

There is nothing unusual in this, and it is consistent with Vercel's commercial strategy. It is still worth stating plainly, so that the open-source argument is not read as a guarantee of full portability today.

Adding an AI agent to an existing Next.js project with eve

This is possible, and the point is underplayed in most presentations of eve, which mostly show projects created from scratch.

In practice, the integration relies on a withEve() function added to the existing Next.js configuration. This function looks for an agent folder in the project and mounts it alongside your application, on the same development server and within the same deployment. There is no need to start again from an empty project or to manage two separate applications.

Another detail matters for hosting outside Vercel. Eve documents a dedicated environment variable that sets the service's public origin when the application is not deployed on Vercel's infrastructure. The Next.js mounting layer is therefore designed to work elsewhere, even though the critical infrastructure layer (sandbox, durability) remains more dependent on the Vercel ecosystem until the alternative adapters mature.

Example of an e-commerce AI agent for customer support

Consider an e-commerce site built with Next.js, with a Medusa backend or an equivalent store. Customer service spends a disproportionate amount of time on repetitive questions about where an order is, how to request a return or when a delivery will arrive. This is precisely the kind of work an eve agent could take on, while the person who handles complex cases stays in charge of them.

What the agent would do in practice

• A get_order_status.ts tool that queries the order database, scoped to the signed-in customer's permissions, to answer directly on the status of an order
• A "returns policy" skill loaded on demand, which gives the agent the context it needs to answer correctly without weighing down every exchange
• An initiate_return.ts tool that creates the return request in the system

Why human approval matters here

This is where human-in-the-loop stops being a theoretical notion. A refund below a certain amount can be processed automatically. Above a defined threshold, the action pauses and waits for approval from a staff member, who has the full conversation trace at hand to make an informed decision. The customer is not left waiting in silence, since the conversation carries on normally while the approval happens in the background.

Customer service therefore stays in place, and human intervention is kept for the cases where it adds real value.

An AI agent to qualify leads on a company website

On a company website, particularly for a service business, the agent can play a different role, which is to qualify an enquiry before it lands in an inbox. A visitor asks about the services on offer, the agent asks about their needs, budget and timeline, then offers a time slot if the profile fits, or passes the request straight to a person if it falls outside the standard scope. Sorting happens before the first human contact instead of after it.

An AI agent to query data on a B2B platform

On an app or internal dashboard built with Next.js, an agent can act as a natural-language interface to your own data. Instead of digging through filters and tables, an employee asks a question directly and gets an answer scoped to their access rights.

• A query_metrics.ts tool that queries the internal database or API, with permissions matched to the signed-in person's role
• A "business definitions" skill that explains to the agent how to calculate a conversion rate or an average order value according to your own rules rather than generic definitions
• An answer in natural language, with a chart generated on the fly when the question lends itself to one

Vercel documents this kind of internal use itself with its d0 agent, which answers more than 30,000 questions a month about the company's data, each query staying scoped to the permissions of the person who asks it. For non-technical staff, the value is that they no longer need to learn a BI tool to get a simple answer.

An AI agent for a shop or service business

On the website of a shop or a service business, an agent can take over part of the time spent on the phone or answering the same questions by email, such as opening hours, availability, bookings or details about a service. Unlike a scripted chatbot that cycles through three canned answers, the agent understands the question asked and replies with the actual context of the business.

• A check_availability.ts tool connected to the calendar, to offer an available slot directly
• A "services and pricing" skill that gives the agent the context it needs to answer without making information up
• Booking handled within the conversation, without sending the visitor to a separate form

For a small business with no one dedicated to answering enquiries throughout the day, this frees up real time otherwise spent on repetitive tasks.

At Nualt

On this kind of subject, we work the same way as we do on hosting or on the choice of a CMS, following the same logic as for our Payload CMS stack. We choose the infrastructure according to the project, and never the other way round.

For a client already hosted on Vercel, adding an eve agent to an existing Next.js application happens within the default setup, with no added complexity. For a client on self-hosted infrastructure, the integration remains possible, but it requires configuring the adapters that replace Vercel's default services, especially for sandboxing generated code. That is genuine integration work, and it is exactly the kind of technical question we investigate before proposing it to a client, instead of passing on a marketing promise as it stands.