Methodology

Where the data comes from

We read each sandbox’s own pricing page, documentation, API reference and GitHub repository, in that order. Every record lists its sources with the date we accessed them, and every number in the table links to the page it came from. A record needs at least three sourced fields beyond its name and summary to be listed.

What counts as an agent sandbox here

An environment an AI agent operates the way a person would: a cloud browser it drives over CDP, Playwright or MCP, or a full desktop (Linux, Windows or macOS) it drives with screenshots, mouse and keyboard. Sandboxes that only execute code are left to sandboxproviders.com and e2balternatives.com; where a code-sandbox vendor also sells a desktop or browser (E2B, Daytona, Hopx), only that product is listed here. Agent products that use a browser but do not sell the browser itself are left out.

Prices

Price per hour uses the lowest published usage rate on any self-serve plan, before plan fees. The basis names the required tier, conversion and traffic exclusions or inclusions. Credit prices are converted only when both unit rates are published. A resource-priced desktop names its published configuration. AgentCore’s full-utilization example is excluded from the minimum-price column. BrowserCat’s subscription comparison excludes optional USDC packs. Bandwidth-only products carry no hourly price.

Models, tools and hosted environments

Google Agent Platform Computer Use provisions hosted browser containers; its preview announcement is dated May 26, 2026. ADK’s computer-use toolset takes a BaseComputer implementation and documents Playwright setup. We count the hosted environment separately from the framework. The former Project Mariner page redirects to DeepMind’s home; it supplies no separately rentable browser offer for this comparison.

OpenAI’s Agents API documents an OpenAI-hosted browser. The Agents SDK / Responses computer-use path also supports developer-supplied drivers and partner environments. CUA/Operator model or consumer-agent access is not counted as an extra infrastructure product.

Anthropic’s Docker reference environment requires your own hosting and a tool loop; it is labelled as a reference environment. Microsoft Foundry Browser Automation uses Playwright Workspaces, so those share one infrastructure row. Azure’s computer-use model tool requires your application to execute actions in your sandbox. Windows 365 for Agents is a separate Cloud PC product.

Solari and AgentBay each have distinct browser and desktop products, listed separately. Counts describe environments, not unique companies.

Limits and features

Max session and Concurrency are the highest values on a self-serve plan; enterprise “custom” tiers are left out. Feature flags (stealth, CAPTCHA solving, live view, recording, CDP, MCP, computer-use API) are set to yes only where a vendor page states the feature, and to no only where a page states it is not offered. An empty cell means we found no statement either way.

How we verify

Each row shows the date we last checked it. When a sandbox does not publish a figure, the cell stays empty.

Corrections

Vendors and readers can send corrections to hello@agentsandboxlist.com with a public source. We update the row and its verified date. Featured placement is paid and labelled; it never changes a row’s data or its position in the sorted table.

Reuse

The dataset is free to reuse under CC BY 4.0 with a link to agentsandboxlist.com. Download it as CSV or JSON.