Kits

A kit packages a set of capabilities a sandbox can use, such as:

• Tools to install
• Environment variables to set
• Credentials to inject
• Domains to allow
• Files to drop in
• Startup commands to run

You declare these in a single spec.yaml file, point the CLI at the directory (or a ZIP, OCI artifact, or Git URL), and the sandbox applies and enforces them at runtime. Credentials stay on the host and go through a proxy instead of entering the VM, and outbound traffic is restricted to the domains the kit allows.

A kit is either a mixin or an agent:

• Mixin kits (kind: mixin) extend an existing agent with extra capabilities. Stack several on the same sandbox.
• Agent kits (kind: agent) define a full agent from scratch: its image, entrypoint, network policies, and everything else the agent needs.

What kits can do

Run commands

A kit can run commands inside the sandbox automatically. Install commands run once at creation; startup commands run each time the sandbox starts.

Install commands are the place to put anything an agent needs into the image, via apt, pip, npm, curl | bash, or whatever fits:

commands:
  install:
    - command: "apt-get update && apt-get install -y jq"

Startup commands cover things like launching background services, warming caches, or refreshing config on each start:

commands:
  startup:
    - command: ["sh", "-c", "my-daemon &"]

Inject files

Kits can inject files into the sandbox in two ways: static files bundled with the kit, and initFiles written at startup with runtime values substituted in.

Static files work well for content that doesn't vary between sandboxes, such as tool configurations, shared linter rules, helper scripts the agent can invoke, or reference material like a style guide or API cheatsheet.

my-kit/
├── spec.yaml
└── files/
    ├── home/
    │   └── .config/my-tool/settings.json
    └── workspace/
        └── .editorconfig

initFiles cover content that depends on runtime values, such as an absolute workspace path that a tool needs to bake into its config file at startup:

commands:
  initFiles:
    - path: /home/agent/.my-tool/config.json
      content: '{"workspace": "${WORKDIR}"}'
      onlyIfMissing: true

See initFiles in the spec reference for all fields.

Set environment variables

Environment variables set by the kit are available to the agent at runtime. Sensitive values can be marked proxy-managed, so the real credential is substituted only when the proxy forwards a request. The secret itself never enters the VM:

environment:
  variables:
    MY_TOOL_WORKSPACE: /home/agent/my-tool
  proxyManaged:
    - MY_TOOL_API_KEY

Control network access

Network rules define which domains the sandbox can reach. For authenticated services, a domain can be mapped to a service identifier, and the proxy injects the auth header on forwarded requests:

network:
  allowedDomains:
    - api.example.com
  serviceDomains:
    api.example.com: my-service
  serviceAuth:
    my-service:
      headerName: Authorization
      valueFormat: "Bearer %s"

Declare credential sources

Credential sources tell the proxy where to find secrets on the host. The sandbox never sees the value itself. The proxy reads it and injects it into outbound requests:

credentials:
  sources:
    my-service:
      env:
        - MY_SERVICE_API_KEY

Define an agent

Agent kits declare an agent: block with the image the agent runs in and the command the user attaches to when they launch the sandbox:

agent:
  image: "my-registry/my-agent:latest"
  entrypoint:
    run: [my-agent, "--yolo"]

See Agent kits for use cases and an example.

Mixin kits

A mixin kit extends an existing agent with extra capabilities. Common use cases:

• Pre-install tools: linters, libraries, or other custom programs
• Grant the agent access to a new authenticated service (a database, a vendor API)
• Inject shared team config (linter rules, editor settings, dotfiles)

Example: Python linting kit

This kit installs Ruff and injects a shared configuration file, so every sandbox starts with the same linting setup.

ruff-lint/
├── spec.yaml
└── files/
    └── workspace/
        └── ruff.toml

schemaVersion: "1"
kind: mixin
name: ruff-lint
displayName: Ruff Linter
description: Python linting with shared team config

network:
  allowedDomains:
    - pypi.org
    - files.pythonhosted.org

commands:
  install:
    - command: "uv tool install ruff@latest"
      user: "1000"
      description: Install Ruff

line-length = 100

[lint]
select = ["E", "F", "I"]

Tip

The templates for the built-in agents (claude, codex, etc) already includes uv, so this mixin can use it without installing it separately.

To run an agent with this mixin:

$ sbx run claude --kit /path/to/ruff-lint/

Agent kits

An agent kit defines a full agent from scratch — image, entrypoint, and everything the agent needs. Common use cases:

• Package a custom agent you've built so others can run it
• Ship a team-internal agent with defaults baked in
• Run a fork of an existing agent with your own config
• Prototype a new agent integration

Agent kits declare everything a mixin kit can, plus an agent: block that tells the sandbox how to launch the agent. For a step-by-step walkthrough, see Build your own agent kit.

Example: the built-in claude agent

The claude agent you get from sbx run claude is defined as a kit. Here is an abbreviated version of its spec, showing how the agent block combines with network, credentials, environment, and commands:

schemaVersion: "1"
kind: agent
name: claude
agent:
  image: "docker/sandbox-templates:claude-code-docker"
  aiFilename: CLAUDE.md
  persistence: persistent
  entrypoint:
    run: [claude, "--dangerously-skip-permissions"]

network:
  serviceDomains:
    api.anthropic.com: anthropic
    console.anthropic.com: anthropic
  serviceAuth:
    anthropic:
      headerName: x-api-key
      valueFormat: "%s"
  allowedDomains:
    - "claude.com:443"

credentials:
  sources:
    anthropic:
      env:
        - ANTHROPIC_API_KEY

environment:
  variables:
    IS_SANDBOX: "1"

commands:
  install:
    - command: "curl -fsSL https://claude.ai/install.sh | bash"
      user: "1000"
      description: Install Claude Code

Using kits

Kits can be loaded from a local path (a directory or ZIP file), a Git repository, or an OCI registry. Pass --kit more than once to stack several kits on the same sandbox.

Important

--kit only takes effect when a sandbox is created. Passing it against an existing sandbox name fails with --kit can only be used when creating a new sandbox. To extend a running sandbox with a kit, use sbx kit add instead.

Local

Point --kit at a directory or ZIP file on disk:

$ sbx run claude --kit ./my-kit/
$ sbx run claude --kit ./my-kit-1.0.zip

While iterating on a kit, apply changes to a running sandbox with sbx kit add instead of recreating it:

$ sbx kit add my-sandbox ./my-kit/

kit add re-runs install commands and re-copies files. Kits can't be removed from a running sandbox — remove and recreate it to start clean.

Git repository

$ sbx run claude --kit "git+https://github.com/docker/sbx-kits-contrib.git#ref=v0.1.0&dir=code-server"

• #ref=<branch|tag|commit> pins to a specific revision. Defaults to the repository's default branch.
• #dir=<path> loads a kit from a subdirectory.
• git+ssh:// URLs also work, using your local SSH agent, Git credential helpers, and .netrc.
• Quote the URL in shells where & starts a background job.

OCI registry

$ sbx run claude --kit ghcr.io/myorg/my-kit:1.0

For Docker Hub, include the full docker.io prefix. See Packaging and distribution for publishing.

Packaging and distribution

The sbx kit subcommands validate, inspect, and publish kits:

• sbx kit validate <path> — check that a kit directory or ZIP is well-formed.
• sbx kit inspect <path> — display kit details. Add --json for machine-readable output.
• sbx kit pack <path> -o <file.zip> — package a directory as a ZIP file for sharing.
• sbx kit push <path> <ref> — publish to an OCI registry (for example, ghcr.io/myorg/my-kit:1.0).
• sbx kit pull <ref> — download a kit from a registry as a ZIP file to the working directory.

For Docker Hub, include the full docker.io prefix — sbx doesn't add it automatically.

Spec reference

A kit directory has a required spec.yaml and an optional files/ tree:

my-kit/
├── spec.yaml       # required
└── files/          # optional — static files to inject
    ├── home/
    └── workspace/

Top-level fields

schemaVersion: "1"
kind: <mixin | agent>
name: <name>
displayName: <name>
description: <text>

The sections below apply to both kinds. Agent kits also declare an agent: block.

Credentials

credentials:
  sources:
    <service-id>:
      env: [<env-var>, ...]
      file:
        path: <path>
        parser: <parser>
      priority: <env-first | file-first>

Service identifiers link credentials to network rules.

Network

network:
  allowedDomains: [<domain>, ...]
  serviceDomains:
    <domain>: <service-id>
  serviceAuth:
    <service-id>:
      headerName: <header>
      valueFormat: <format>

Environment

environment:
  variables:
    <NAME>: <value>
  proxyManaged: [<NAME>, ...]

Variable names must be valid shell identifiers ([A-Za-z_][A-Za-z0-9_]*).

Commands

commands:
  install:
    - command: <shell-string>
      user: <uid>
      description: <text>
  startup:
    - command: [<argv>, ...]
      user: <uid>
      background: <true | false>
      description: <text>
  initFiles:
    - path: <path>
      content: <text>
      mode: <octal>
      onlyIfMissing: <true | false>
      description: <text>

install

Runs once during sandbox creation. Shell strings passed to sh -c.

startup

Runs at every sandbox start. String array, not interpreted by a shell.

Startup commands are non-interactive. They run before the agent attaches, with no terminal connected, so they can't prompt the user (for example, an interactive aws login will hang or fail). They also don't gate the agent's entrypoint: the agent launches once startup commands have been dispatched, regardless of background. Use them for non-interactive prep — launching daemons, warming caches, refreshing config — and use commands.initFiles for any value that needs to land on disk before the agent runs.

initFiles

Files written at sandbox start, with runtime substitution.

Static files

my-kit/files/
├── home/       → /home/agent/
└── workspace/  → primary workspace path

Parent directories are created automatically. Existing files are overwritten. Absolute paths and path-traversal sequences (../../) are rejected.

Agent block

Required for kind: agent.

agent:
  image: <image-ref>
  aiFilename: <filename>
  persistence: <persistent | ephemeral>
  entrypoint:
    run: [<argv>, ...]
    args: [<arg>, ...]

Base image requirements

The agent's container image must provide:

• A non-root agent user at UID 1000 with passwordless sudo.
• A /home/agent/ home directory owned by agent.
• HTTP proxy environment variables (HTTP_PROXY, HTTPS_PROXY, NO_PROXY) preserved across sudo.
• The agent binary (baked in, or installed via commands.install).

Build on top of docker/sandbox-templates:shell-docker to get these for free.

Memory

memory: |
  <markdown>

Top-level field. Markdown appended to the agent's memory file at sandbox creation. The agent reads this content at startup, so write it as instructions or notes the agent should follow when working in the sandbox. Applied only when agent.aiFilename is set.

The file is written to the parent of the workspace path inside the sandbox, not to the workspace itself. For a workspace mounted at /Users/you/myproject, the memory file lands at /Users/you/AGENTS.md (or whatever aiFilename is set to). It exists only inside the sandbox — nothing is written to the host.

Debugging

When a kit doesn't behave as expected, two commands cover most cases:

• sbx policy log shows every outbound request the sandbox proxy saw, the rule it matched, and how it was forwarded (forward-bypass, forward, block). It's the primary tool for diagnosing install-time download failures, blocked domains, or unintended TLS interception. It was the diagnostic that surfaced the serviceDomains-too-wide install corruption documented in Build your own agent kit.
• sbx exec <sandbox> -- <cmd> runs an arbitrary command inside an existing sandbox. Useful for inspecting post-install state without recreating: which mytool, ls /home/agent/.local/bin/, cat /home/agent/.config/..., and so on.

Install and startup command output isn't currently retained after sandbox creation — it scrolls past in the sbx run / sbx create output and isn't captured anywhere accessible afterward. For now, the two commands listed plus a clean recreate (sbx rm <sandbox> && sbx run …) are the practical loop.