I found figma’s remote MCP to be a poor fit for iOS development (it sets tokens on fire and gives Claude the details as React+Tailwind) and got much better results by having Claude build a set of “lens” tools around their REST API (“give me all of the fonts”, “give me the layout dimensions”, “give me the colors”, etc).
The key to making it token efficient was allowing Claude to invent its own plaintext markup format for the lens output.
We don’t know what quantization level was used for the weights or the kv cache for you or for parent poster, so this is probably an apples to oranges comparison.
The KV cache is probably offloaded to RAM after a generation is complete. Then it is pulled back into any rack in the data center that has the model you’re using.