The Agent Skills Directory

[CREDENTIALS_UNSAFE]: The skill's documentation and tool definitions for publishing agents and prompts (post_add-agent, post_add-prompt) explicitly mention accepting a private_key argument for blockchain tokenization. Requesting or passing raw private keys within an AI agent's tool arguments is a dangerous practice that exposes users to potential credential theft or unauthorized financial transactions.
[EXTERNAL_DOWNLOADS]: In the 'Connecting' section for Claude Desktop, the skill instructs users to run npx -y mcp-remote. This command dynamically downloads and executes the mcp-remote package from the npm registry at runtime.
[COMMAND_EXECUTION]: The Python integration example in Section 3 references the httpx2 library. This appears to be a typo for the standard httpx library, but following this instruction could lead users to install a potentially unvetted or malicious package from PyPI.
[INDIRECT_PROMPT_INJECTION]: The skill implements several tools that fetch user-generated content from an external marketplace, creating an attack surface where third-party listings could influence the agent's behavior.
Ingestion points: Tools such as get_get-agents, get_get-prompts, and get_reviews in SKILL.md retrieve external content from swarms.world.
Boundary markers: The instructions provide no delimiters or 'ignore embedded instructions' warnings for the agent when processing this remote data.
Capability inventory: The skill allows the agent to perform network write operations (publishing/editing) to the vendor's API.
Sanitization: No sanitization, escaping, or validation of the marketplace content is required before it is returned to the agent context.

swarms-marketplace-mcp