IronClaw Docker-in-Docker

Thin wrapper on nearaidev/ironclaw that adds dockerd so IronClaw can run sandboxed workloads in nested containers. Sysbox is the preferred runtime; --privileged can be used as a less-isolated fallback when Sysbox is unavailable or unsupported.

What it adds

• Docker CLI/daemon binaries + iptables
• A DinD entrypoint that starts dockerd, verifies the baked worker image is available in the inner Docker store, then execs ironclaw

Everything else (the ironclaw binary, runtime, user) comes from the base image.

Build

The build helper pulls the sandbox worker image at build time, saves it to an archive, loads it into a temporary DinD container's inner Docker store, and commits the final image. Runtime startup then uses the local inner-Docker image instead of pulling from the network.

bash scripts/build-dind-image.sh \
  ironclaw-dind \
  nearaidev/ironclaw:latest \
  local

By default the baked sandbox image is derived from the IronClaw image tag (nearaidev/ironclaw:<tag> -> nearaidev/ironclaw-worker:<tag>). Override it with:

SANDBOX_IMAGE=nearaidev/ironclaw-worker:<tag> \
bash scripts/build-dind-image.sh ironclaw-dind nearaidev/ironclaw:<tag> <source-digest>

The bake step starts a temporary DinD container with --runtime=sysbox-runc by default. On hosts where Sysbox is unavailable or unsupported, set DIND_BAKE_RUN_ARGS=--privileged as a less-isolated compatibility fallback.

Run

docker run --runtime=sysbox-runc ironclaw-dind

The entrypoint fails fast if the sandbox image is missing and does not let IronClaw auto-pull it by default. To restore the old runtime-pull fallback for development, set SANDBOX_PULL_IF_MISSING=true.

Do not mount /var/lib/docker as a persistent volume for normal deployments. That directory contains the baked inner-Docker image store; persisting it across image upgrades can mask the worker image bundled in the new ironclaw-dind image.