daytona-sandbox

Daytona Agent Skills for Claude

A Claude Agent Skill that teaches Claude how to drive Daytona — secure, elastic sandboxes for running AI-generated code. It is the procedural-knowledge layer that sits on top of Daytona's existing connectivity (the Python/TypeScript SDK, the daytona CLI, and the daytona-mcp server).

Why a Skill (and not just the MCP)?

The Daytona MCP server is the connection — 12 thin, stateless tools over Daytona's API. A Skill is the know-how: when to spin up a sandbox, how to chain a build when exec is stateless, when to snapshot, how to expose a preview, and to always clean up. They're complementary layers, per Anthropic's own guidance:

"If you're explaining how to do something, that's a skill. If you need Claude to access something, that's MCP."

The Skill also drives the Python SDK, which is a strict superset of the MCP's 12 tools (sessions, code_run, snapshots, volumes, full git, rich filesystem, lifecycle, computer-use, LSP), and it costs ~100 tokens of context when idle versus a dozen tool schemas loaded into every conversation. See skills/daytona-sandbox/reference/sdk-vs-mcp.md for the full map.

Launch post & diagram

A short launch post, "Skills vs MCPs," lives in blog/ (source · compiled PDF). Its central diagram — the "capability cake," where MCP and Skills are different layers, not competitors — is available standalone in blog/diagram/ (TikZ source + PDF + PNG):

Layout

.
├── .claude-plugin/
│   └── marketplace.json          # plugin registry (publish this repo as a marketplace)
└── skills/
    └── daytona-sandbox/
        ├── SKILL.md              # entrypoint: overview + lifecycle playbook + links
        ├── LICENSE.txt
        ├── reference/            # loaded on demand (progressive disclosure)
        │   ├── sandboxes.md      # create / configure / snapshot / volumes / destroy
        │   ├── exec.md           # commands, code_run, sessions, env, git
        │   ├── files.md          # filesystem operations
        │   ├── preview.md        # preview links
        │   └── sdk-vs-mcp.md     # what the MCP covers vs the SDK superset
        ├── examples/
        │   ├── run-python.md
        │   └── clone-and-test.md
        └── scripts/              # executed via bash; code never enters context
            ├── healthcheck.py    # verify DAYTONA_API_KEY + connectivity
            └── run_in_sandbox.py # deterministic create -> run -> destroy helper

Prerequisites

• A Daytona API key: set DAYTONA_API_KEY in the environment (get one at https://app.daytona.io), or run daytona login.
• Python with the SDK for the bundled scripts: pip install daytona.

Install

Claude Code (via plugin marketplace)

/plugin marketplace add daytonaio/claude-skills      # this repo
/plugin install daytona-sandbox@daytona-skills

Claude Code (local, for development/testing)

Copy or symlink the skill into a skills directory Claude Code reads:

# project-scoped (this repo only):
mkdir -p .claude/skills && cp -r skills/daytona-sandbox .claude/skills/
# or personal (all projects):
cp -r skills/daytona-sandbox ~/.claude/skills/

Then verify it's picked up with /skills, or invoke directly with /daytona-sandbox.

claude.ai / Claude API

• claude.ai: zip the skills/daytona-sandbox folder and upload via Settings → Capabilities → Skills.
• API: upload via the /v1/skills endpoints (workspace-wide).

Note: Skills don't sync across surfaces — install on each surface you use. Bundled scripts need a code-execution environment with network access (Claude Code has it; the bare Claude API surface does not — there, use the daytona-mcp server instead).

Verify

python3 skills/daytona-sandbox/scripts/healthcheck.py

License

Apache-2.0. See skills/daytona-sandbox/LICENSE.txt.