Summary and Next Steps

What you built

In this track you went from running agents directly to packaging sandbox environments as declarative, shareable kits.

What you did	How
Shipped a Claude Code skill into the sandbox	Mixin kit with files/workspace/ injection
Understood proxy-managed credential injection	credentials.sources + serviceAuth
Forked the built-in claude agent	Agent kit with custom entrypoint
Stacked two kits on one sandbox	--kit --kit
Pulled a kit from the community repo	git+https://github.com/docker/sbx-kits-contrib.git#dir=…

---

The files you created

kits/
├── docker-review/
│   ├── spec.yaml
│   └── files/
│       └── workspace/
│           └── .claude/skills/docker-review/SKILL.md
└── claude-safe/
    └── spec.yaml

These are real, working kits. Commit them to this repo and anyone can use them with a single --kit flag.

---

What kits give you that shell scripts don't

• Credentials never enter the VM - the proxy reads them from the host and injects per-request
• Composable egress rules - stack kits without mutating a shared allowlist file
• Shareable via Git URL or OCI registry - no clone, no setup doc, no config dance
• Validated before run - sbx kit validate catches spec errors before they waste a sandbox creation
• TCK-testable - write a Go test file using tck.NewSuiteFromDir and CI validates your kit against a real container

---

Next steps

• Browse the community kits: github.com/docker/sbx-kits-contrib
• Read the full spec reference: docs.docker.com/ai/sandboxes/customize/kits
• Build an agent kit from scratch: docs.docker.com/ai/sandboxes/customize/build-an-agent
• Contribute a kit to the community repo - the CONTRIBUTING.md explains the process