MCP-Agent

The agent dynamically executes shell commands via `StdioTransport` (e.g., `npm install`, `npx`) based on server configurations read from `config.json` or generated by the LLM. If a malicious MCP server configuration is added (manually by the user or through a compromised prompt), it could lead to arbitrary code execution on the host machine. The agent also connects to external MCP servers, meaning a malicious server could exploit vulnerabilities in the client or provide unsafe tool outputs. While `ast.literal_eval` is used for action inputs, `json.loads` is also present, which doesn't protect against malicious data if the structure is not enforced. There's no explicit sandboxing mentioned for tool execution or subprocess calls.