OpenRouter adds hosted code execution across multiple AI models

OpenRouter has introduced two beta tools that let AI models run commands in a provider-hosted sandbox, including models from different companies. The practical draw is that developers can add code execution without building and maintaining the sandbox themselves, while retaining a route across supported models.

OpenRouter Watch analysis

What happened

OpenRouter’s guide describes openrouter:shell, available on its Responses and Messages APIs, and openrouter:bash, available on Messages. Both run commands in OpenRouter’s sandbox when configured to do so; the tools are in beta, so their APIs may change. They work through OpenRouter’s global endpoint, not its in-region endpoints or Chat Completions API.

The sandbox is an isolated container scoped to an account and workspace. Outbound network access is off by default, and OpenRouter says execution costs $0.0001 per second, with a 30-second minimum for a new or sleeping container. A request can make up to 30 server-tool calls. The company’s comparison of hosted code-execution tools sets those details alongside offerings from OpenAI, Anthropic and Google.

Why it matters

Hosted execution can spare a team the work of provisioning, patching and securing a container for short jobs such as running a script or checking a file. OpenRouter’s distinction is that its tools sit at the routing layer, so they can be used with models supported by the relevant APIs rather than only one company’s models.

That convenience has boundaries. The guide says the sandbox suits short, bounded work; a custom base image, GPU workload or session lasting hours still calls for infrastructure the developer operates. And because the tools are in beta, teams should expect details to change.

Our read

This is a useful piece of plumbing for developers who want a model to run a bounded command without building a sandbox first. The cross-model angle is the point, not a claim that every model or workflow will behave identically. Try it on a small, non-sensitive task, check the current limits and costs, and keep the work that needs specialised environments in a sandbox you control. Fewer containers to babysit is a fine improvement, even if the babysitting was wearing a very technical hat.

What to watch

• Whether the beta tools’ API, pricing or execution limits change.
• Which models and workflows work reliably through each supported API.
• Whether OpenRouter publishes further detail on sandbox isolation and the tools’ operational limits.

Discussion spark: For short AI-agent tasks, would you rather use a provider-hosted sandbox across models or operate your own for greater control?

not affiliated with or endorsed by OpenRouter