Source: Shell tool guide (OpenAI docs)
OpenAI's shell tool in hosted (managed Debian container, network off by default, org allowlist plus per-request network_policy, sidecar-injected domain secrets) and local modes.
Contents
- What: official API documentation
- Publisher: OpenAI
- Date: living document, read 2026-11-14
- Original: Shell tool
Summary
The Responses API shell tool runs commands either in OpenAI-hosted containers or in a runtime you host. The hosted mode is a general-purpose container (several languages, skills mounted as bundles) with explicit egress controls and secret substitution outside the model’s view.
Key claims
- Hosted environments:
container_auto(provisioned per request) orcontainer_reference(a long-running container from the Containers API with a memory limit and inactivity expiry). Responses API only. - Image: Debian 12; working directory
/mnt/data; no TTY, nosudo; Python 3.11, Node.js 22.16, Java 17, PHP 8.2, Ruby 3.1, Go 1.23 preinstalled. - Network: no outbound access by default. Enabling it takes an org-level admin allowlist plus a request-level
network_policythat can only narrow it; requests naming other domains fail. domain_secretsbind a named secret to an allowed domain; the model sees only a placeholder such as$API_KEYand a sidecar substitutes the real value for approved destinations only.- Skills are versioned bundles mounted into the container.
- Local mode returns
shell_callitems for your own executor (replaces the retiredlocal_shelltool); OpenAI recommends command allowlists, approvals for destructive actions, sandboxing with limits, own timeouts and full logging. - Network-fetched content may carry prompt injection; review logs of requested and actual destinations.