For your agent

Agents learn SafeClaw from one file: safeclaw-skill.md. It is the single source of truth for agent behavior: discovery, phantom placement, sc run, approvals. This page is only about handing it over.

The easy way

The console's "Connect a new agent" install prompt already pairs the machine, mints the agent's env, and delivers the skill. If you went through Quickstart step 3, you are done; this page is for doing it by hand.

Paste-ready prompt

Machine already paired, agent env already minted (sc agent add ran)? Send your agent this:

Fetch https://raw.githubusercontent.com/SafeClaw-OSS/safeclaw/main/static/safeclaw-skill.md
and save it where you keep always-loaded skills. Follow it whenever a task
needs an external credential: use the phantom placeholders it describes and
route those commands through `sc run --`. Never ask me to paste a key and
never try to read one; if a service isn't connected yet, give me the console
link as the skill instructs.

Agents with filesystem access can equally copy the file from a checkout of this repo into their skills directory.

What the agent's env means

sc agent add <name> mints one env block per agent:

Both lines are non-secret, so the env block is safe to display. The vault is not part of the env: each request names one in its URL (/v/<vault>/…), and sc run binds one per launched command (sc run --vault <id> -- <cmd> to pick). sc vault ls lists ids (* = default).

One identity per agent keeps the audit trail attributable; don't share an identity file across agents, and don't re-run sc agent add when the env already exists (it registers a duplicate).

These vars steer the agent's requests only. Your own sc commands read the active vault from ~/.safeclaw/config.toml (set by sc login / sc vault use), so a stale agent env can never hijack what you do at the terminal.