T3 Code with Claude Opus 5.5: builds and voice agent tests

This livestream follows development of Clear Mud HQ, a centralized workspace for the speaker's web apps. T3 Code is the main coding assistant harness, with Claude Opus 5.5 handling development and Claude Sonnet 5.5 tested through game prototypes. Sonnet initially fails to appear in the model selector; later, the speaker opens its generated games and describes them as functional but visually limited.

Much of the work concerns a voice AI agent inside HQ. The speaker wants to keep a conversation open, mute it while dictating elsewhere, and ask it to update the content calendar. Early requests go to the wrong page or become code changes requiring preview and approval. After further development, the agent reports creating and replacing livestream entries, but the YouTube script workflow still encounters missing API access and timeouts.

The speaker explores computer use and a sandboxed Linux VM on the hosting server as possible ways to support more actions. This is a server-hosted workspace experiment using external model services; the stream does not demonstrate local LLM inference. Other changes include shared notes, simpler video modules, and plans for invitation-based Portal workspaces with separate client data. For business evaluations, the speaker argues for testing actual workflows and pain points rather than relying on visual demos.