The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill facilitates the creation of tools that interact with web page content, which is an inherent surface for indirect prompt injection attacks. The documentation proactively addresses this by requiring developers to use untrustedContentHint annotations, implement strict schema validation (additionalProperties: false), and adhere to specific risk classes for tool preparation and commitment.
[EXTERNAL_DOWNLOADS]: The skill references well-known external tools for browser automation and validation, specifically the browse CLI and the Stagehand library (@browserbasehq/stagehand). These resources are used for their intended development purposes and are hosted on standard public registries.
[REMOTE_CODE_EXECUTION]: Instructions provide links to external source workspaces for tool compilers and generators (e.g., webmcp-gen, add-webmcp) hosted under the Browserbase organization on GitHub. These are presented as trusted upstream components for the WebMCP integration workflow.
[CREDENTIALS_UNSAFE]: The documentation explicitly prohibits the hardcoding or return of secrets, including tokens, cookies, or internal stack traces. It also provides guidance to skip scanning sensitive files such as .env and credentials during the application inventory process.

webmcp