Why we're building Cusp

Coding agents changed how a lot of us work over the past year, and almost all of them are tied to one lab's models. Meanwhile the models keep trading places: the strongest model for a hard refactor changes every few months, open-weight models have become good and much cheaper, and the same model can score very differently depending on the agent that runs it. We think the agent should be the constant and the model should be a choice you can change mid-session. So we're building Cusp: one agent with the tools, permissions, skills, subagents, hooks and MCP support we rely on every day, working with any model or gateway, tuned to each model family, and running on one runtime across a desktop app, a CLI and, later, the web. There isn't much to show yet. We'll write here as it comes together, and we'll be plain about what works and what doesn't.