The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill instructs users to install pinned versions of official vendor packages from the @1claw scope on npm, such as @1claw/mcp and @1claw/cli. These are standard dependencies for the 1Claw service functionality.
[COMMAND_EXECUTION]: Includes a utility script scripts/validate-setup.sh that uses curl and jq to verify API connectivity and credential validity against the 1Claw backend. The script enforces HTTPS and restricts requests to trusted hostnames by default.
[DATA_EXFILTRATION]: The skill transmits the agent's API key (ocv_ prefix) to the official 1Claw API (api.1claw.xyz) solely for the purpose of exchanging it for a short-lived authentication token, which is the intended security architecture of the service.
[PROMPT_INJECTION]: While static analysis flagged the string 'ignore previous instructions', this occurs within a defensive context. The skill provides explicit guidelines and a dedicated inspect_content tool to help the agent identify and ignore malicious instructions embedded in untrusted data.
[INDIRECT_PROMPT_INJECTION]:
Ingestion points: Untrusted data enters the agent context through get_secret tool outputs, external API responses, and content passed to the inspect_content scanner (SKILL.md, references/mcp-and-api.md).
Boundary markers: The instructions explicitly command the agent to treat instructions found in untrusted content as data and follow strict human-in-the-loop confirmation policies for sensitive actions.
Capability inventory: The skill possesses capabilities for blockchain transaction signing (submit_transaction), secret modification (put_secret), and executing shell scripts for environment validation (SKILL.md, scripts/validate-setup.sh).
Sanitization: Employs a dedicated inspect_content MCP tool designed to detect injection, social engineering, and PII in arbitrary text strings before further processing.

1claw