In early 2026, there's a debate on how to structure tool descriptions for large language models (LLMs). One approach focuses on creating top-level tools for specific tasks, like listing pull requests on GitHub. Another method leverages existing shell commands, allowing the model to generate command pipelines instead of making direct tool calls. This second approach could reduce costs and improve efficiency by enabling the model to execute a series of commands without needing to re-prompt for each step. The author shares experiences with integrating Google Docs and Groups into an agent's workflow. For Google Docs, they utilized Google’s API to create a command-line interface (CLI) that lists and exports documents. However, this felt redundant since many SaaS providers have OpenAPI specs that could streamline this process. They discovered Restish, a tool that interprets OpenAPI specs, simplifying the interaction with APIs. Although Restish manages API authorization, the author prefers to handle that independently, leading them to develop a wrapper script that accommodates their needs. For OAuth 2.0 authorization, the author found success with the oauth2c command-line client, which simplifies the process of obtaining and refreshing tokens. This reduced a complex Python script into a concise shell script. They also tackled secure token storage on macOS using the Keychain. By using the security command with specific flags, they managed to store sensitive tokens securely, requiring biometrics for access. This setup allows for seamless token handling while minimizing the risk of exposure. The article also touches on the challenge of accessing less developer-friendly services, like Google Groups, which lack an official API for exporting discussion histories. This highlights the ongoing need for better interoperability between tools and services, especially those that do not provide machine-readable specifications.