The Agent Skills Directory

[DYNAMIC_EXECUTION]: The example code for a custom calculator tool utilizes the Function constructor to evaluate mathematical expressions: Function("'use strict'; return (${expression})")(). This pattern allows for the execution of arbitrary JavaScript code from the user-provided expression string.\n- [COMMAND_EXECUTION]: The SDK documentation details the Bash and BashSession tools, which enable the agent to execute shell commands on the host system. The examples demonstrate how to use these tools in conjunction with the bypassPermissions configuration.\n- [INDIRECT_PROMPT_INJECTION]: The SDK establishes an attack surface for indirect prompt injection by facilitating the processing of untrusted data (via file reads, web fetches, and user prompts) and providing examples where the agent is configured to bypassPermissions. This configuration allows the agent to use sensitive tools like Bash, Write, and Delete without human confirmation.\n
Ingestion points: Data enters the agent context through agent.prompt(), query(), and the results of tools like Read, Glob, and http_get.\n
Boundary markers: The provided examples do not demonstrate the use of delimiters or specific safety instructions to isolate data from instructions.\n
Capability inventory: The SDK includes a suite of tools for file management, process execution (Bash), and network operations.\n
Sanitization: There is no demonstration of sanitization or validation for input data processed by the example tools.\n- [EXTERNAL_DOWNLOADS]: The documentation includes examples that use npx to download and execute Model Context Protocol (MCP) servers, such as @modelcontextprotocol/server-filesystem and @modelcontextprotocol/server-postgres.

open-agent-sdk-typescript