Vercel eve Puts Your AI Agents in Your Own Codebase

Vercel open-sourced eve on June 17 at Ship 26 in London. It is a TypeScript agent framework under Apache 2.0, it shipped in public beta, and the company was already running more than 100 agents on it internally before anyone outside got to look at it.

I wrote separately about Hyperagent, which is the platform answer for an operator who wants agents doing work without building anything. eve is the other half of that decision. It is what you reach for once a workflow has proven itself and you want it living in your own codebase, on infrastructure you control, next to your own data.

The reason that matters to someone who has no intention of ever writing TypeScript is narrower than the launch coverage suggested, and more useful. eve changes the price of the boring parts, and the boring parts are what have been killing custom agent projects at companies your size.

An Agent Is a Directory

The design decision underneath eve is that an agent is a folder on disk, and the folder structure is the configuration. There is no registration step and no wiring file.

A minimal agent has agent.ts, which sets the model through Vercel's AI Gateway, and instructions.md, which is the system prompt written in plain Markdown. From there you add capability by adding files:

• tools/ holds TypeScript files. Drop in get_weather.ts exporting a defineTool and the model can call get_weather. The filename becomes the tool name.
• skills/ holds Markdown files. These are playbooks the agent loads on demand rather than carrying in context permanently.
• subagents/ holds directories. A subagent is the same shape one level down, with its own instructions, its own tools, and its own sandbox, and the parent calls it the way it calls a tool.

You scaffold the whole thing with npx eve@latest init my-agent. Once it is running, the agent is an HTTP service. You POST a message to /eve/v1/session to start a durable session, then attach to /eve/v1/session/<id>/stream to watch lifecycle events come back.

If the vocabulary sounds familiar, it should. Instructions, tools, skills, subagents, and scheduled tasks are the same primitives showing up across every serious agent system right now, including the one I use to write and ship this site. eve's contribution is making them filesystem conventions that a build step discovers, which means the diff for "give the agent a new capability" is one new file.

The Six Things You Stop Building

This is the actual argument, and it has nothing to do with syntax.

Every custom agent project I have watched stall out died in the same place. The agent worked in a demo on someone's laptop in an afternoon. Then it needed to survive a model timeout halfway through a run, execute untrusted commands without touching anything else, stop and wait for a human before it sent something, report what it did in a way an auditor could read, and prove it was still correct after a prompt change. That second list is where the months go, and none of it is the thing you wanted.

eve ships six of those out of the box:

That list is the whole pitch. A framework that gives you a nice way to call a model saves you a week. A framework that gives you durable execution, isolation, approvals, and evals saves you the quarter where the project usually gets quietly shelved.

Tracing deserves a specific note, because it is the one operators underrate. eve records a span hierarchy per turn, down to individual tool calls, and Vercel's MCP server exposes get_agent_run_trace so you can pull the exact reasoning, tool inputs, and token usage for a production run. When someone asks why the agent emailed a client the wrong number, that is the difference between an answer and a shrug.

What Vercel Runs On It

Vercel published its own internal fleet alongside the launch, which is more useful than a feature list because these are agents with owners and budgets.

d0 is a data analyst living in Slack. Anyone at the company asks it a question and it answers from the warehouse, handling more than 30,000 questions a month, which makes it their most-used internal tool. Every query is scoped to the asker's own permissions, so it cannot show you a table you were not already allowed to see. That permission-scoping detail is the part worth stealing regardless of what you build on.

Lead Agent works inbound leads the moment they arrive and follows up on its own. Vercel says it costs about $5,000 a year to run and returns 32 times that, maintained part-time by one engineer.

Vertex handles support and resolves 92% of incoming tickets without escalation.

Athena answers pipeline and forecast questions against Snowflake and Salesforce in plain language. It was built by the RevOps team in 6 weeks without engineers.

V sits in front of everything in Slack and routes each request to whichever agent in the fleet should handle it.

Treat the return figures as what they are, which is a vendor reporting on itself with no third party checking the math. The shapes are still the useful part. Four of those five are the same four workloads I pointed at in the Hyperagent piece, which were internal data questions, inbound lead work, support triage, and recurring reporting. Two companies with completely different stacks arrived at the same starting lineup, which tells you something about where this technology is actually ready.

The routing agent is the one I would flag for anyone planning past their first agent. Once you have more than three, something has to decide which one gets the request, and Vercel solved that with another agent rather than a menu.

What It Costs

eve itself is free and open source. You pay for what it runs on, which is Vercel Functions for compute, Workflows for durable sessions, Sandbox for isolated execution, and AI Gateway or your own provider for the model calls.

Sandbox is the line worth understanding, because it is priced on active CPU. You are billed $0.128 per vCPU-hour only while your code is actually executing, plus $0.0212 per GB-hour of memory for the full wall-clock duration. Time spent waiting on model inference, a database query, or an API response does not bill CPU. For agent workloads, which are almost entirely waiting, Vercel puts the saving at up to 95% against conventional per-second billing. Sandbox creation runs $0.60 per million, and egress is billable.

AI Gateway passes provider list rates through with no markup and includes $5 a month of credits on the free tier.

The honest read is that costs are legible here in a way they are not on a consumption-priced platform. You can look at a trace, see the tokens and the CPU time, and know what a run cost and why. What you give up is the platform's willingness to absorb the operational work, and an engineer's time is more expensive than either line item.

Where It Locks You In

The license is Apache 2.0 and the framework uses an adapter pattern, so porting is permitted. Permitted is doing a lot of work in that sentence.

Durable execution, sandboxed compute, and credential brokering each resolve to a Vercel service. Moving eve to another runtime means finding or writing a replacement for each one. The adapter boundary makes that technically possible and does not make it pleasant. If you have a multi-cloud requirement or you need to self-host, eve is a poor fit today, and it is a poor fit by design rather than by omission. This is the Next.js pattern running again, and how you feel about that is probably already settled.

Public beta means public beta. Practitioners working with it early hit canary AI SDK builds that broke a tool loop with a type-validation error mid-run, and a too-new Connect release that shipped a verifier rejecting valid events. Pin your versions and commit the lockfile. The API will move under you.

Two smaller things worth knowing before you scope anything. Observability is thin for failures that happen before a run starts, so an event that never lands is harder to see than a run that goes wrong. And schedules tighter than once a day need a paid plan, which matters if you assumed hourly was free.

Vercel also shipped Passport at the same event, a governance layer for what they call shadow AI, meaning the internal tools your engineers already built on personal accounts and shared credentials. It is in public beta. If you have more than a handful of developers, that inventory problem is real whether or not you ever touch eve.

Platform First, Then Framework

The decision between this and a platform is not close, and it goes in a specific order.

Pilot on a platform. Prove the workflow produces output a human agrees with, learn what the prompt actually needs to say, and find out whether anyone uses the thing. That costs you a few hundred dollars and 30 days, and most candidate workflows die there, which is the point.

Build on eve when the workflow has survived that and one of three things is true. The agent needs to live inside your product rather than beside it. Compliance says the data cannot leave your environment. Or your volume has grown to where per-run platform pricing costs more than running it yourself.

Building first is how companies end up with a beautifully engineered agent nobody asked for. I have watched that happen more than once, and the engineering was never the problem.

Deterministic automation still wins everything it was already winning. If the steps never vary, a scenario in Make or n8n is cheaper and more reliable than any agent, and our Make vs n8n vs Claude Code comparison covers where that line falls. Agents earn their cost on the work that used to need a person to look at it and decide. For the broader taxonomy, the autonomous AI agents guide covers how these categories fit together, and the MCP servers guide covers the protocol most of these tools use to reach your systems.

One number from Ship is worth carrying out of all this. Vercel said the share of deployments triggered by coding agents went from under 3% to more than half in roughly 6 months. Whatever you think of the framework, that is the environment your competitors are operating in.

What We Do With This

Most of our work sits exactly on the transition this post describes. A client proves something on a platform, it works, and then it needs to run inside their own systems against their own data with their own controls on it. eve is a real option for that second step now, and it was not six months ago.

It does not change the thing that actually determines whether any of this works. An agent amplifies the process you already have. Point it at a clean pipeline and it compounds. Point it at a mess and you get the mess faster and with better logging. The teams that get value here are the ones who fixed the process first and then automated it, in that order.

If you want a second opinion on which of your workflows belong on a platform, which belong in your own stack, and which should stay a Zapier scenario nobody touches, get in touch.

Frequently Asked Questions

What is Vercel eve? eve is an open-source TypeScript agent framework Vercel released on June 17, 2026 under Apache 2.0. An agent is a directory of files: agent.ts for model config, instructions.md for the system prompt, and folders for tools, skills, and subagents that the framework discovers automatically at build time.

Is eve free? The framework is free and open source. You pay for the infrastructure it runs on, which is Vercel Functions, Workflows, Sandbox, and model calls through AI Gateway or your own provider.

How is eve different from Hyperagent? eve is a framework developers build on, and the agent lives in your codebase and your infrastructure. Hyperagent is a hosted platform an operator configures by talking to it. Most companies should prove a workflow on a platform first and move it onto a framework once it has earned the engineering time.

Can I run eve outside Vercel? The Apache 2.0 license permits it and the adapter pattern makes it technically possible, but durable execution, sandboxed compute, and credential brokering each map to a Vercel service you would need to replace. Teams with multi-cloud or self-hosting requirements should treat eve as a poor fit for now.

Is eve ready for production? Vercel runs more than 100 agents on it internally, including the support and sales agents above, so it is production-ready in their hands. It is also in public beta, and early users have hit breaking changes from unpinned dependencies. Pin your versions and commit the lockfile.

What should the first agent do? Internal research or recurring reporting. Both produce output someone on your team reviews before it reaches a customer, so a bad run costs you an hour rather than a relationship.