Add BrowseCLI in a Vercel Sandbox example

solutions/browsecli-vercel-sandbox/.env.example

@@ -0,0 +1,26 @@
+# --- Browserbase (the cloud browser the agent drives) ---
+# Get a key at https://www.browserbase.com/settings
+BROWSERBASE_API_KEY=bb_live_xxxxxxxxxxxxxxxxxxxxxxxx
+
+# --- Vercel Sandbox auth (to create the Firecracker microVM) ---
+# Token: https://vercel.com/account/tokens
+# Team + Project IDs: Vercel dashboard → Project → Settings
+VERCEL_TOKEN=xxxxxxxxxxxxxxxxxxxxxxxxxxxx
+VERCEL_TEAM_ID=team_xxxxxxxxxxxxxxxxxxxxxxxx
+VERCEL_PROJECT_ID=prj_xxxxxxxxxxxxxxxxxxxxxxxx
+
+# --- Model auth: Vercel AI Gateway (no separate Anthropic key needed) ---
+# The agent's model (anthropic/claude-sonnet-5) is served through Vercel AI
+# Gateway, so one Vercel auth covers both the Sandbox and the model. Pick ONE:
+#
+# 1. OIDC (recommended, keyless): run `vercel link` then `vercel env pull
+#    .env.local` in this directory. Vercel writes a short-lived
+#    VERCEL_OIDC_TOKEN to .env.local that the Gateway reads automatically — no
+#    extra key to manage. (Refresh it with `vercel env pull` when it expires.)
+#
+# 2. AI Gateway API key: create one at https://vercel.com/dashboard/ai-gateway
+#    and set it here. Use this if you prefer a long-lived key over OIDC.
+# AI_GATEWAY_API_KEY=vck_xxxxxxxxxxxxxxxxxxxxxxxx
+
+# --- Optional: override the agent's task ---
+# TASK=Find the latest Node.js LTS version and its release date from nodejs.org.

solutions/browsecli-vercel-sandbox/.gitignore

@@ -0,0 +1,37 @@
+# See https://help.github.com/articles/ignoring-files/ for more about ignoring files.
+
+# Dependencies
+/node_modules
+/.pnp
+.pnp.js
+
+# Next.js
+/.next/
+/out/
+next-env.d.ts
+
+# Production
+build
+dist
+
+# Misc
+.DS_Store
+*.pem
+
+# Debug
+npm-debug.log*
+yarn-debug.log*
+yarn-error.log*
+
+# Local ENV files
+.env
+.env.local
+.env.development.local
+.env.test.local
+.env.production.local
+
+# Vercel
+.vercel
+
+# typescript
+*.tsbuildinfo

solutions/browsecli-vercel-sandbox/README.md

@@ -0,0 +1,50 @@
+# Browser agent with bash-tool in a Vercel Sandbox
+
+A [Vercel AI SDK](https://ai-sdk.dev) agent that uses Vercel's [`bash-tool`](https://github.com/vercel-labs/bash-tool) to run the [`browse`](https://www.npmjs.com/package/browse) CLI inside a [Vercel Sandbox](https://vercel.com/docs/vercel-sandbox) — an ephemeral Firecracker microVM. The `browse` CLI drives a real browser running on [Browserbase](https://www.browserbase.com), so the browser itself never runs in the microVM.
+
+## How it works
+
+```
+┌──────────────────────────────┐        ┌──────────────────────────────┐     CDP/wss   ┌────────────────────────────┐
+│  Host (your machine / CI)    │        │  Vercel Sandbox (microVM)    │  ──────────▶  │  Browserbase cloud browser │
+│                              │        │  node24                      │               │                            │
+│  ToolLoopAgent (AI SDK)      │ ─bash─▶ │  browse CLI runs here        │ ◀──────────── │  navigates / reads pages   │
+│  reasoning + tool loop       │        │  (installed by sandbox.ts)   │   page data   │                            │
+└──────────────────────────────┘        └──────────────────────────────┘               └────────────────────────────┘
+```
+
+- The **agent loop runs on the host**. `sandbox.ts` provisions the sandbox, installs the `browse` CLI inside it, builds a `bash` tool with `bash-tool` (`createBashTool({ sandbox })`), and runs a `ToolLoopAgent`.
+- The **model is served through [Vercel AI Gateway](https://vercel.com/docs/ai-gateway)**. A bare `provider/model` string (`anthropic/claude-sonnet-5`) resolves through the default Gateway provider in `ai@6`, authenticated by Vercel's OIDC token — so the same Vercel auth that powers the Sandbox also powers the model, and no `ANTHROPIC_API_KEY` is needed.
+- Only the **`bash` tool executes inside the sandbox**. The model navigates the web by emitting `browse` commands, which `bash-tool` runs in the microVM.
+
+The default task is a product-research example: on Amazon, search for the current top mechanical keyboards and, for the top 5 results, compare each product's title, price, star rating, and number of ratings — returning a comparison table with each product's URL. It 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. The agent plans its own steps — there are no site-specific instructions in the prompt. Override the goal with the `TASK` env var.
+
+## Setup
+
+The agent's model (`anthropic/claude-sonnet-5`) is served through [Vercel AI Gateway](https://vercel.com/docs/ai-gateway), so **one Vercel auth covers both the Sandbox and the model — there is no separate `ANTHROPIC_API_KEY`.**
+
+You need:
+
+- `BROWSERBASE_API_KEY` — the cloud browser ([browserbase.com/settings](https://www.browserbase.com/settings)). It is passed to the sandbox as a default environment variable (stored encrypted by Vercel and injected into the `browse` commands), so it is never written to a file in the sandbox.
+- Vercel Sandbox credentials — `VERCEL_TOKEN` ([vercel.com/account/tokens](https://vercel.com/account/tokens)), `VERCEL_TEAM_ID`, and `VERCEL_PROJECT_ID` (from your Vercel project settings).
+- Model auth via AI Gateway — pick one:
+  - **OIDC (recommended, keyless):** link this directory to a Vercel project and pull its env. The `vercel` CLI writes a short-lived `VERCEL_OIDC_TOKEN` that the Gateway reads automatically — nothing else to manage:
+    ```bash
+    vercel link                # link to your Vercel project
+    vercel env pull .env.local # writes VERCEL_OIDC_TOKEN
+    ```
+  - **AI Gateway API key:** create one at [vercel.com/dashboard/ai-gateway](https://vercel.com/dashboard/ai-gateway) and set `AI_GATEWAY_API_KEY` in `.env`. Use this if you prefer a long-lived key over OIDC.
+
+```bash
+pnpm install
+cp .env.example .env       # fill in BROWSERBASE_API_KEY + Vercel Sandbox creds
+vercel link                # OIDC path: link the project…
+vercel env pull .env.local # …and pull VERCEL_OIDC_TOKEN (skip if using AI_GATEWAY_API_KEY)
+pnpm start
+```
+
+> AI Gateway requires the Vercel team to have a payment method / billing enabled to service model requests (free credits unlock once a card is on file). This is the only model-side prerequisite — there is no Anthropic account or key.
+
+A successful run provisions a sandbox, prints each `browse` command the agent issues through the `bash` tool, and ends with a `===== FINAL ANSWER =====` summary.
+
+The default uses a plain `--remote` Browserbase session, which works on any plan. Verified browsers (residential IP + automatic CAPTCHA solving) are a Scale-plan upgrade — see [browserbase.com/pricing](https://www.browserbase.com/pricing).

solutions/browsecli-vercel-sandbox/package.json

@@ -0,0 +1,24 @@
+{
+  "name": "browsecli-vercel-sandbox",
+  "version": "1.0.0",
+  "private": true,
+  "license": "MIT",
+  "repository": "https://github.com/vercel/examples.git",
+  "description": "A bash-tool agent that runs the browse CLI inside a Vercel Sandbox, driving a Browserbase cloud browser.",
+  "type": "module",
+  "scripts": {
+    "start": "tsx sandbox.ts",
+    "typecheck": "tsc --noEmit"
+  },
+  "dependencies": {
+    "@vercel/sandbox": "^2.2.1",
+    "ai": "^6.0.0",
+    "bash-tool": "^1.3.0",
+    "just-bash": "^3.0.1"
+  },
+  "devDependencies": {
+    "@types/node": "^20.14.0",
+    "tsx": "^4.19.0",
+    "typescript": "^5.5.0"
+  }
+}