Quickstart

This page explains how to install SafeClaw, run agent-driven setup, approve the browser steps, and make the first brokered call.

1. Install

curl -fsSL https://safeclaw.pro/install.sh | sh

The installer uses no sudo, verifies the download's SHA256, puts sc in ~/.local/bin, and installs the skill at skills/safeclaw/SKILL.md. When run by an agent, it continues straight into sc setup; in a terminal it prints one next step. It also installs the skill with sc skill install. Official sources are safeclaw.pro and github.com/getsafeclaw/safeclaw. Supported platforms: Platform support.

2. Run setup

Run this as the person or as their agent:

sc setup

--name <a short name for the agent> is optional and defaults to the agent's product name, such as claude-code or codex. sc setup never prompts. Each run prints what it needs from the person, if anything, and the command to run next:

To pair this machine, open this link and enter the code:
  https://app.safeclaw.pro/pair
  Code: 4821-0937-5563 (expires in about 15 minutes)
Next: sc setup --name codex --wait

The agent gives the person the link and code, runs Next:, and repeats until a run prints no Next:. After an error it runs the printed Fix:. The last Next: is usually sc unlock, which prints an unlock link and an sc op wait command. Continue to the next section after approval. With --json, each run prints one object whose status is action_required, waiting, done or error, with the same values as fields.

In the browser, the person signs in or creates an account the first time, creates a passkey, and approves the machine. The passkey also creates the first vault. They then approve the agent and choose which vaults it may use. Under the hood, setup runs sc login if the machine is not paired and sc agent add if the agent is not registered. It writes ~/.safeclaw/agents/<name>/agent.env and installs the skill. It is safe to run again.

At a terminal, a person can still use sc login to pair a machine or sc agent add <name> to add an agent by hand, such as a second agent. sc setup runs whichever of those is still needed and also handles the env file and skill.

3. Add a credential

Two doors, same result: a connection with hosts and a phantom.

Console (recommended for OAuth services like GitHub or Gmail): open your vault's Connections tab, pick the service, sign in or paste the key. Values you type there are encrypted in your browser before upload.

Terminal, for any bare API key:

sc set HF_TOKEN --host huggingface.co

Prompts for the value, asks for one passkey approval, and mints the phantom __sc__hf_token__ anchored to huggingface.co. Entered this way the value never leaves your machine.

4. Unlock the vault

After approving the machine and the agent, run sc unlock. It prints an unlock link for the person and the sc op wait command. For several vaults, run sc vault ls, choose one, then run sc unlock --vault <name>.

5. First brokered call

Put the phantom where the credential belongs and route the command through sc run --:

HF_TOKEN=__sc__hf_token__ sc run -- huggingface-cli whoami
GITHUB_TOKEN=__sc__github__ sc run -- gh pr list

The proxy swaps the phantom for the real value on the way out, only toward that connection's own hosts. The command gets the API response; the environment, the shell history, and the agent hold only the phantom.

6. When an approval is needed

Policy-gated uses don't fail silently: the command's error output carries an approval link. Open it, tap your passkey, re-run the same command. Agents can block on sc op wait <op_id> instead of asking you to say "done". Approvals are cached per policy, so routine calls don't nag.

7. Daily rhythm

sc up      # start of day: daemon up + vault unlocked, one passkey tap
sc lock    # done for the day: wipe keys from the daemon's memory
sc doctor  # something's off: health + reachability checks

Next: teach the agent the full patterns in sc run and phantoms, or hand it the skill directly via For your agent.