The Agent Skills Directory

[REMOTE_CODE_EXECUTION]: The Aerodrome plugin (plugins/aerodrome.md) utilizes uvx to download and execute the sugar-sdk from its official GitHub repository (github.com/velodrome-finance/sugar-sdk) to facilitate liquidity and swap operations.
[REMOTE_CODE_EXECUTION]: The Morpho plugin (plugins/morpho.md) is designed to run the @morpho-org/cli package via npx to prepare and simulate lending transactions.
[CREDENTIALS_UNSAFE]: The Uniswap plugin (plugins/uniswap.md) contains a hardcoded x-api-key (NeoYO3V50_koJAipDEalYWbMO1XMaFPAQmpOm6_Npo0) used for authenticating requests to its trade API gateway.
[COMMAND_EXECUTION]: The approval flow reference (references/approval-mode.md) instructs the agent to execute shell commands (open, xdg-open, or start) to automatically launch the user's browser for transaction approvals when running in CLI-enabled environments.
[EXTERNAL_DOWNLOADS]: The skill includes a fallback mechanism in SKILL.md that fetches instruction and reference files from https://docs.base.org/ai-agents/skills/ using the web_request tool if local filesystem access is unavailable.
[EXTERNAL_DOWNLOADS]: Multiple plugins perform network operations to external DeFi domains (including avantisfi.com, bankr.bot, moonwell.fi, and uniswap.org) to retrieve market data, user positions, and prepared transaction calldata.
[PROMPT_INJECTION]: The skill processes untrusted data from external DeFi APIs, creating a surface for indirect prompt injection.
Ingestion points: Responses from web_request calls to various protocol APIs across all plugin files.
Boundary markers: No specific delimiters or "ignore instructions" warnings are defined for the ingested data.
Capability inventory: The skill utilizes sensitive tools like send_calls and sign to execute on-chain actions based on the fetched data.
Sanitization: The instructions include steps to normalize CLI output and require the agent to review simulation results and transaction summaries before presenting approval links to the user.

base-mcp