The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The tool ingests content from local project files, including AGENTS.md and project-specific SKILL.md files, to provide context for the AI agent. This data ingestion is paired with high-privilege capabilities such as shell script execution and system interaction via MCP, creating a surface for indirect prompt injection.\n
Ingestion points: The skill reads SKILL.md files from ~/.agents/skills/ and ./.agents/skills/, and project context from AGENTS.md in the current working directory.\n
Boundary markers: The documentation does not describe explicit markers or instructions to isolate untrusted project data from the agent's core instructions.\n
Capability inventory: The tool can execute user-defined shell scripts (notify and webSearchTool configurations), run remote MCP servers with filesystem/database access, and execute slash commands.\n
Sanitization: There is no evidence of sanitization or filtering of the content retrieved from these project files before it is processed by the agent context.\n- [EXTERNAL_DOWNLOADS]: The skill requires the global installation of the @vegamo/deepcode-cli package from the NPM registry. It also utilizes npx to fetch and execute various Model Context Protocol servers (e.g., @modelcontextprotocol/server-github) during runtime as part of its integrated features.\n- [REMOTE_CODE_EXECUTION]: The configuration for MCP servers leverages npx -y to download and run code at execution time. While these packages originate from the official Model Context Protocol organization, the pattern involves dynamic remote code execution based on the tool's configuration settings.\n- [COMMAND_EXECUTION]: The assistant is designed to trigger arbitrary shell scripts for notifications and web search functionality. Example configurations include the use of curl within these scripts to transmit data to external webhooks, which are executed in the local user environment.

deepcode-cli