The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill performs several global npm installations. While packages from Anthropic (@anthropic-ai/claude-agent-acp) and OpenAI (@openai/codex) originate from trusted organizations, the skill also installs 'opencode' from a non-verified source.
[REMOTE_CODE_EXECUTION]: The skill uses npx @skills-hub-ai/cli install to download and execute additional software components. The '@skills-hub-ai' organization is not a verified or well-known service, posing a risk of executing untrusted code in the user's environment.
[COMMAND_EXECUTION]: The skill executes multiple shell commands to detect editor versions, list extensions, and register agents via CLI tools like devin agent add. This involves high-level access to the development environment configuration.
[INDIRECT_PROMPT_INJECTION]: The skill contains logic to read preferred skill names from local project files (.skills-hub.json or CLAUDE.md) and pass them directly to an installation command (npx @skills-hub-ai/cli install). This creates a vulnerability where malicious instructions or package names in these project files could lead to unauthorized software installation.
Ingestion points: Reads content from .skills-hub.json and CLAUDE.md in the project root.
Boundary markers: None. Content from these files is used directly as arguments.
Capability inventory: Subprocess execution via npm, npx, and editor CLIs; file system writes to ~/.devin/agents.json, .zed/agents.json, etc.
Sanitization: None; the skill uses the names found in the files directly to fetch and install packages.

acp-agent-setup