My old gaming GPU now handles 3 jobs I used to pay Claude for

Claude, Anthropic's AI service, charges a flat $20 per month regardless of usage. This can be problematic for users who want to budget their expenses based on actual usage, as excess usage can result in additional charges. The author of this account, however, found an alternative solution by utilizing the processing power of their gaming GPU to run local AI models. Their RTX 3070 graphics card, originally intended for gaming, now handles a range of tasks that would otherwise be outsourced to Claude.

The author emphasizes the importance of keeping private documents, such as health tests, insurance paperwork, and bank statements, secure and off cloud AI data centers. By running local AI models, they can maintain privacy and avoid potential issues with cloud services. They specifically recommend LM Studio, a local AI application, for uploading files and conducting tasks.

LM Studio allows users to attach various file types directly into a chat, and if a file is too large to fit into the model's context window, it will extract relevant portions for the user's query. The author also mentions the Filesystem MCP feature, which enables the model to open files from chosen folders on the user's machine.

To keep browsing private, the author uses Jan, a free, open-source app for running local models, which comes with its own browser agent. Jan is installed via a Chrome extension and connects to a local model via the Model MCP Server. Site access is set to "On Click" and file URL access is disabled, ensuring that browsing remains private.

The author notes that while newer Claude models offer better token efficiency, this doesn't necessarily mean users will be more careful with their usage. They continue to monitor their daily and weekly usage of Claude and only delegate simple tasks to their local models, preserving Claude for more complex work. Their preferred local model for chat is Qwythos 9B v2, while Empero AI's fine-tune of Qwen 3.5 9B is used for faster, less intensive tasks.

By dividing their workload between Claude and local models, the author is able to stay within their budget while maintaining privacy and security.