The Agent Skills Directory

[EXTERNAL_DOWNLOADS]: The skill facilitates the installation of the native-devtools-mcp package from NPM and provides instructions to clone and build a Rust binary from a GitHub repository (sh3ll3x3c/native-devtools-mcp).
[REMOTE_CODE_EXECUTION]: Instructs users to use npx -y to download and immediately execute the MCP server. It also includes steps for building the executable from source using cargo build, which involves downloading and executing build-time dependencies.
[COMMAND_EXECUTION]: Provides capabilities to launch native applications with arbitrary command-line arguments, control system input (keyboard and mouse), and execute arbitrary JavaScript expressions within a browser context using the cdp_eval tool.
[DATA_EXFILTRATION]: Captures potentially sensitive information through full-screen and window-specific screenshots, OCR-based text extraction, accessibility tree snapshots, and DOM snapshots from browsers.
[PROMPT_INJECTION]: The skill is vulnerable to indirect prompt injection due to its processing of untrusted data from the user interface and the web.
Ingestion points: Screen content via take_screenshot, find_text, and take_ax_snapshot; browser DOM content via cdp_take_dom_snapshot (File: SKILL.md).
Boundary markers: None identified in the instructions or tool documentation to help the agent distinguish between its own logic and instructions embedded in processed UI data.
Capability inventory: Significant capabilities including launch_app, type_text, click, and cdp_eval which can be misused if an attacker influences the agent via on-screen content (File: SKILL.md).
Sanitization: The skill does not describe any validation or sanitization of content extracted via OCR or the accessibility tree.

native-devtools-mcp-automation