The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill processes data from external web pages using functions like view-screen and get-deck. This untrusted content could potentially contain instructions aimed at influencing the agent's behavior.
Ingestion points: Web content is ingested from the browser tab in SKILL.md using page-level tools.
Boundary markers: The instructions do not specify delimiters to separate untrusted tool output from agent instructions.
Capability inventory: The skill can execute JavaScript (javascript_tool, Runtime.evaluate), navigate the browser, and perform data mutations using the page's own tools.
Sanitization: There is no mention of sanitizing or validating the content retrieved from the browser before processing.
[DYNAMIC_EXECUTION]: The skill instructs the agent to execute JavaScript within the page's context to interface with the window.__agentNativeWebMcp helper or document.modelContext API. This is accomplished using javascript_tool (Claude Code) or Runtime.evaluate via CDP (Codex).
[COMMAND_EXECUTION]: The README.md includes a command to install the skill: npx @agent-native/skills@latest add --skill webmcp. This involves the execution of a remote package from the NPM registry as part of the setup process.

webmcp