Agent skills
Beachside ships three agent skills from its repository: beachside and beachside-cli in skills/, and add-provider in .agents/skills/. A skill is a SKILL.md file that an agent loads when a task needs it, so the guidance stays out of your prompt until it is useful.
| Skill | Teaches an agent to |
|---|---|
beachside |
Write application code with beachside and a provider adapter: connect, create and clean up sandboxes, run commands, move files, and handle uncertain outcomes without repeating work. |
beachside-cli |
Create, use and remove sandboxes from the shell with the beachside command, read its JSON output and exit codes, and clean up after failures. |
add-provider |
Write a provider adapter with @beachside/adapter: verify the account, map the shared setup options, implement create and destroy, test with adapterSuite, then package and document it. See Write an adapter. |
Install
Section titled “Install”Install with the skills installer, which supports Claude Code, Codex, Cursor and other agents:
npx skills add pandemicsyn/beachside --skill beachside -gnpx skills add pandemicsyn/beachside --skill beachside-cli -gnpx skills add pandemicsyn/beachside --skill add-provider-g installs the skill for your user. Leave it out to install into the current project instead, so everyone working in the repository gets it; that suits add-provider, which belongs in the adapter’s repository.
The CLI also carries its own copy of beachside-cli, matching its version. beachside skill install writes it to ~/.agents/skills (and ~/.claude/skills when Claude Code is set up), and beachside skill install --project writes it into the current folder. beachside skill prints it.
For an agent without a skill system, copy the SKILL.md contents into its project or user instructions.
What the skills do not do
Section titled “What the skills do not do”The skills never hold credentials. The SDK skill reads keys from environment variables; the CLI skill uses your existing beachside.json profiles and tells the agent to leave configuration to you. Both tell the agent that live sandboxes cost money, to clean up what it creates, and never to retry an operation whose outcome is unknown.
For one-off tasks you can still paste the SDK prompt or the provider integration prompt instead.