The Agent Skills Directory

[INDIRECT_PROMPT_INJECTION]: The skill describes tools that ingest external or untrusted data, creating a potential surface for indirect prompt injection.
Ingestion points: Tools such as mcp/meshchatx fetch data from GitHub issues, and mcp/agents reads repository documentation and markdown files.
Boundary markers: The skill explicitly documents security constraints such as path jailing (restricting file access to allowed roots) and secret redaction.
Capability inventory: The toolkit includes capabilities for file system reads, repository build commands (make), and authenticated GitHub API operations (gated by environment tokens).
Sanitization: The instructions mandate the use of redaction for sensitive material and the resolution of all paths to a jailed root to prevent directory traversal.
[COMMAND_EXECUTION]: The skill provides instructions for running standard development commands like make all, make test, and gosec. These are localized to the developer's environment for building and auditing the toolkit and are consistent with the skill's stated purpose as a developer resource.
[DATA_EXFILTRATION]: While the skill mentions environment variables like GITHUB_TOKEN and reading local configuration in ~/.reticulum, it explicitly mandates that secret material must be redacted and never returned by the tools, mitigating exfiltration risks.

mcp-toolkit