selfware

Defines a unified file protocol bundling data, logic, and views in one unit, enabling decentralized human-agent collaboration with canonical data authority.

This skill is a protocol specification (draft) and is instruction-only, so it poses minimal direct installation risk. Things to check before using or trusting a runtime that implements this protocol: 1) verify the concrete implementation's source (homepage/repo) and review its code for network calls and file I/O; 2) confirm the runtime actually binds to localhost by default and enforces the 'confirmation_required' rules (no silent apply, rollback points) before fetching or applying remote canonical data; 3) be cautious when enabling remote canonical sources or granting an implementation permission to 'send context' or contact external endpoints — those actions can leak local data if not properly gated; 4) prefer implementations with an auditable codebase and reproducible builds. Overall the skill is internally coherent with its stated purpose, but safety depends on any runtime implementation you choose to run.

Type: OpenClaw Skill Name: selfware Version: 1.0.0 The SKILL.md file defines a complex 'Selfware Protocol' that instructs the agent to perform high-risk behaviors, including sending 'partial context' (logs, state, and intent) to external discovery endpoints (e.g., floatboat.ai) and fetching/applying remote updates. While the protocol documentation emphasizes user confirmation and 'No Silent Apply' policies, it establishes a framework for data exfiltration and remote code execution. The instructions are highly prescriptive and redefine the agent's operational boundaries, creating a significant attack surface through external dependencies and remote collaboration features.