From CLI Experiments to Agentic OS: Building Bhindi

We started with a simple CLI-based setup we called Miniature Bhindi, allowing us to trigger personal agents via Cursor. This early experiment helped us realize how lightweight agent deployment could be—users just needed a CLI file and access tokens to perform tasks. No LLMs were involved at this point, apart from Cursor.

As usage grew, we built a basic server with a few agents and a frontend for internal use. Users interacted with agents using access tokens—this proved to be a better alternative to MCPs. Eventually, the number of agents and use cases expanded, and Miniature Bhindi evolved into Bhindi.

Today, Bhindi is a full agentic operating system. A user request triggers a server call, which makes an AI (LLM) call to determine the required tool calls. Bhindi selects or infers the right agent(s), makes the necessary API calls, manages OAuth token flows, and provides intelligent feedback loops if something is missing.

To optimize performance and cost, we built Subagents—breaking tool executions into smaller context windows to avoid hallucination and reduce LLM token usage. We also implemented Backrooms, a communication layer for agents to coordinate or echo tasks to others, and added prompt caching and pricing optimizations.