What

Adds solutions/browsecli-vercel-sandbox — a Vercel AI SDK agent that runs the browse CLI inside a Vercel Sandbox (an ephemeral Firecracker microVM) using Vercel's bash-tool.

The agent loop runs on the host (ToolLoopAgent from ai). Its model (anthropic/claude-sonnet-5) is served through Vercel AI Gateway, and its bash tool — built with createBashTool({ sandbox }) — executes commands inside the sandbox, where the browse CLI is installed. The model navigates the web by emitting browse commands; the browser itself runs remotely on Browserbase, so it never runs in the microVM.

┌──────────────────────────────┐        ┌──────────────────────────────┐   CDP/wss   ┌────────────────────────────┐
│  Host (your machine / CI)    │        │  Vercel Sandbox (microVM)    │  ────────▶  │  Browserbase cloud browser │
│  ToolLoopAgent (AI SDK)      │ ─bash─▶ │  browse CLI runs here        │ ◀────────── │  navigates / reads pages   │
└──────────────────────────────┘        └──────────────────────────────┘  page data  └────────────────────────────┘

System prompt: small, self-discovering

The system prompt is intentionally minimal. It does not enumerate browse subcommands or flags — it tells the agent the CLI is installed and pre-configured, points it at browse --help, and lets it learn the surface itself:

You are an autonomous deep-research agent. You have a `browse` CLI (Browserbase
browser automation) in your bash tool — it is installed, and its auth and a
shared browser session are already configured via environment variables. Learn
how to use it by running `browse --help` (and `browse <command> --help` as
needed), then complete the task. When you cite a document, link the direct
document itself, not a viewer, preview, or index page that wraps it. Return a
clear, well-sourced answer.

This keeps the example honest about how a capable model actually uses a CLI (read --help, then go) and avoids a brittle wall of command cheatsheet text. The one piece of steering that makes the short prompt work is moved out of the prompt and into the environment — see below.

Env-steering for a flag-free shared session

The sandbox is created with two default env vars:

env: { BROWSERBASE_API_KEY, BROWSE_SESSION: 'agent' }

• BROWSERBASE_API_KEY — authenticates browse against Browserbase.
• BROWSE_SESSION=agent — steers browse to run remotely and share one named browser session across commands. Because of this, the agent can just run browse open <url> (no --remote, no --session) and every command lands in the same cloud browser. This is what lets the prompt drop all the session/flag instructions.

Model via Vercel AI Gateway (no Anthropic key)

The model is routed through Vercel AI Gateway rather than a direct Anthropic provider:

• sandbox.ts passes a bare provider/model string — model: 'anthropic/claude-sonnet-5' — which ai@6 resolves through the default Gateway provider.
• The @ai-sdk/anthropic dependency and the ANTHROPIC_API_KEY env var are dropped entirely.
• Gateway reads its credentials from the environment automatically. The recommended path is keyless via OIDC: the Vercel CLI's vercel env pull writes a short-lived VERCEL_OIDC_TOKEN that the Gateway uses — so one Vercel auth covers both the Sandbox and the model. An AI_GATEWAY_API_KEY is supported as an alternative for those who prefer a long-lived key.

This removes a credential from the example: a user who is already authenticated with Vercel (to create the Sandbox) does not need a separate Anthropic account or key.

Default task

A product-research example, with no site-specific instructions in the prompt — the agent plans its own steps. On Amazon it searches for the current top mechanical keyboards and, for the top 5 results, compares each product's title, price, star rating, and number of ratings, returning a comparison table that includes each product's URL. This is a genuinely browser-only task: Amazon search renders its results client-side and returns nothing useful to a plain HTTP fetch, so the agent has to drive a real browser to get the data. Override the goal with the TASK env var.

Using Amazon (https://www.amazon.com), research the current top mechanical
keyboards: search the site, then for the top 5 results compare each product's
title, price, star rating, and number of ratings. Return a comparison table
including each product's URL.

Contents

• sandbox.ts — standalone, directly-runnable script: Sandbox.create({ runtime: 'node24' }) → install browse in the sandbox → build the bash tool → run a ToolLoopAgent (model via AI Gateway), streaming each issued browse command and ending with a ===== FINAL ANSWER ===== summary.
• README.md / package.json / pnpm-lock.yaml / tsconfig.json / .env.example / .gitignore

The example uses pnpm to match the repo convention (packageManager: pnpm@9.13.0; every active solutions/ example commits a pnpm-lock.yaml). The committed pnpm-lock.yaml is generated with pnpm install; install/run commands in the README are pnpm install / pnpm start.

Required env

• BROWSERBASE_API_KEY — the cloud browser.
• VERCEL_TOKEN / VERCEL_TEAM_ID / VERCEL_PROJECT_ID — Vercel Sandbox auth.
• Model auth via AI Gateway — either VERCEL_OIDC_TOKEN (keyless; vercel link + vercel env pull .env.local) or AI_GATEWAY_API_KEY. No ANTHROPIC_API_KEY.

E2E Test Matrix

Run against the exact committed code (sandbox.ts @ this branch) on real Vercel cloud: a freshly provisioned Firecracker microVM, browse CLI installed in-sandbox, ToolLoopAgent loop on the host, model anthropic/claude-sonnet-5 served through AI Gateway. The model run env was asserted at preflight to contain no ANTHROPIC_API_KEY and no AI_GATEWAY_API_KEY — only VERCEL_OIDC_TOKEN (keyless OIDC).

Note: AI Gateway requires the Vercel team to have billing / a payment method on file to service model requests. The keyless OIDC verification above was run on a billing-enabled team; the auth mechanism itself (OIDC → Gateway) is what the change exercises.

Type of Change

• New Example
• Example updates (Bug fixes, new features, etc.)
• Other (changes to the codebase, but not to examples)

New Example Checklist

• 🛫 npm run new-example was used to create the example
• 📚 The template wasn't used but I carefully read the Adding a new example steps and implemented them in the example (MIT license, .env.example + setup instructions, README, .gitignore; uses pnpm with a committed pnpm-lock.yaml per repo convention)
• 📱 Is it responsive? Are mobile and tablets considered? (N/A — this is a CLI/agent script with no UI)

This PR is intentionally a draft. Opening for early visibility / format review before marking ready.