Vercel · The AI builder's cloud
Four Pieces.One AI Stack.
AI Gateway, the AI SDK, Agents, and Sandboxes—Vercel's primitives for putting real AI inside real products. The pieces are powerful and well-documented. Turning them into something you ship and maintain is the skill. That's where we come in.
The Stack
Four primitives, one toolkit
Each piece does one job well. Together they're everything you need to build, run, and trust AI in production—without stitching ten vendors together.
AI Gateway
One key, every model
A single endpoint and API key in front of hundreds of models—OpenAI, Anthropic, Google, and more. Automatic failover when a provider blips, load balancing, and spend tracking, with no markup on tokens.
Swap models with one line and never get locked to a single vendor.
AI SDK
The TypeScript toolkit
The open-source toolkit that's become the default way to build AI in TypeScript—one unified API for text, streaming, tool calls, and structured output, with drop-in hooks for React, Vue, and Svelte.
Put real AI features inside the app you already run.
Agents
Loops that finish the job
The agent loop plus Vercel's workflow primitives: agents that call tools, make decisions, and keep going across steps—durable enough to retry, resume, and survive a crash instead of dying on a timeout.
Go from a one-shot prompt to a system that does multi-step work.
Sandboxes
Run AI-written code safely
Ephemeral, isolated Linux microVMs that execute untrusted or agent-generated code—real shell and filesystem, with the network locked down so nothing leaks. Spin one up, run the code, throw it away.
Let an agent write and run code without risking your machine or data.
How They Compose
One loop, end to end
The four pieces aren't a menu—they're a pipeline. Here's how a real agent comes together when you wire them in order.
Build it with the AI SDK
Text, streaming, tool calls, and structured output—the agent's brain, written once in TypeScript and portable across providers.
Route every call through the Gateway
One key reaches hundreds of models. When a provider slows or fails, traffic fails over automatically—and you can see exactly what each call costs.
Make it durable with Agents
The loop retries, resumes, and holds state across steps, so a long-running task survives a hiccup instead of starting from scratch.
Run its code in a Sandbox
When the agent writes code, it executes in an isolated microVM with the network locked down—safe to run, trivial to discard.
Four primitives, one loop. Knowing which piece each problem needs is where we coach.
Where Thinking Backward Comes In
The demo is easy.
Production is the skill.
A weekend project that calls one model is a different thing from an AI feature your business leans on—one with fallbacks, durability, and a safe place to run code. The gap between them is judgment, not syntax.
We build the first agent alongside you on Vercel's stack, then teach you the patterns so your team keeps shipping its own. Capability that stays with you.
Hands-on, with you
We build the first agent alongside you—and the stack is yours to keep.
Outcome first
We start from the AI feature you actually need and work backward to the smallest stack that ships it.
Built to trust
Failover, durability, and sandboxing baked in—so what you ship holds up in production.
What We Build Together
AI features you own and can extend
We start from the outcome—a real AI capability your product needs—and work backward to the smallest slice of the stack that ships it.
A model strategy on the Gateway
One key, the right model for each job, automatic fallbacks, and spend you can actually see—so you're never hostage to a single provider or a sudden price hike.
A real AI feature in your app
We wire the AI SDK into your own product—streaming, tool calls, structured output—so it's a capability you own, not a chatbot bolted on the side.
An agent that does the work
A durable, tool-using agent for a job that matters to you—built to retry and resume, with the judgment calls and guardrails mapped out.
A safe place to run it
Sandboxed execution for any code your agent writes, with the network locked down—so automation never quietly becomes a liability.