Antigravity Usage Intelligence: Track Tokens, Quotas, Cache, Reasoning, and Agent Activity

Antigravity Usage Intelligence: Track Tokens, Quotas, Cache, Reasoning, and Agent Activity I’ve been using Antigravity heavily for development, and one thing kept bothering me: I could see the agent doing a lot of work, but it was hard to understand how much was actually being used. How many tokens were going through my sessions? How much of that was fresh input versus cached context? How much reasoning was happening? Which projects were using the most? How often was I hitting the quota? And what did my usage actually look like over time? That led me to build Antigravity Usage Intelligence , an open-source usage tracker and analytics dashboard designed around those questions. GitHub: https://github.com/Nir-Bhay/antigravity-usage-intelligence What Antigravity usage actually looks like Antigravity has become much more agent-oriented. With Antigravity 2.0, Google introduced a standalone desktop experience alongside features such as subagents, scheduled tasks, agent management, CLI support, and other tools for longer and more autonomous workflows. That changes the way we think about usage. A simple prompt counter is not enough anymore. An agent can perform multiple turns, call tools, read files, generate output, use reasoning, and reuse cached context during a single task. The official Antigravity documentation also explains that quota limits are tied to the amount of work performed by the agent, which can vary significantly from one task to another. That makes visibility much more useful than simply knowing whether a request succeeded. So what does this project track? The dashboard breaks usage into several parts instead of showing one mysterious number. Token breakdown The dashboard separates: Fresh input tokens Cached prompt tokens Output tokens Reasoning or thinking tokens That distinction matters. For example, two agent sessions can both look like large workloads, while their actual token behavior is very different because one relies heavily on cached context. The