The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: Documentation identifies a recommended installation method that pipes a remote script from the official domain directly to the shell: curl -fsSL https://opencode.ai/install | bash.\n- [EXTERNAL_DOWNLOADS]: The skill mentions multiple ways to install software and extensions, including global npm packages (opencode-ai), Homebrew taps (anomalyco/tap/opencode), and remote MCP servers (e.g., from mcp.sentry.dev).\n- [COMMAND_EXECUTION]: Describes the use of a bash tool for executing shell commands and a shortcut for the agent to execute system commands directly by prefixing messages with !.\n- [DATA_EXFILTRATION]: Documents the existence of webfetch and websearch tools, which allow the agent to perform network requests to external domains and ingest data from the internet.\n- [PROMPT_INJECTION]: Documentation highlights an attack surface for indirect prompt injection via the webfetch and websearch tools.\n
Ingestion points: Untrusted content can enter the agent context through the webfetch and websearch tools as described in reference/tools.md.\n
Boundary markers: The documentation does not define specific boundary markers for external content, though it details a permission configuration system.\n
Capability inventory: The agent has access to powerful tools including shell execution (bash), file creation (write), and file modification (edit, patch).\n
Sanitization: A granular permission system featuring allow, ask, and deny settings for various tools and commands is documented to manage capability risks.

opencode