tai42 lean sandbox session image: sh, python3, and coreutils, nothing else.
10K+
The lean shell session image of the tai42 platform: sh, python3, and coreutils on a slim base, on a non-root user with /workspace as the working directory. Nothing else: no SDK, no platform code, no credentials. It is the runtime behind the langchain_deep_agent sandbox shell.
ghcr.io/tai42ai/tai-sandbox-execThe image is not run by hand. The tai42-sandbox-docker provider starts one container per conversation from it, mounts that conversation's workspace volume at /workspace, and injects the session credentials into the container environment at start.
Point the agent at the image by digest; a bare tag is refused:
TAI_AGENTS_LANGCHAIN_DEEP_SESSION_IMAGE=docker.io/tai42/tai-sandbox-exec@sha256:<digest>
Distribution semver — X.Y.Z, X.Y, X, and latest — cut on the same release as the platform image. Every manifest list is multi-arch (amd64 + arm64), trivy-gated before push, and cosign-keyless-signed in both registries.
Apache-2.0. The full source of the image build lives in tai42ai/tai-distribution.
Content type
Image
Digest
sha256:c286347b0…
Size
41.5 MB
Last updated
about 21 hours ago
docker pull tai42/tai-sandbox-exec