Every small AI app seems to inherit a surprisingly large to-do list.
Connect a model. Store credentials. Add tool access. Collect context. Handle permissions. Then, finally, build the thing the app is for.
With Switchboard, I’m exploring a different starting point.
Build the interface and agent behaviour for a specific job. Connect them to the user’s existing AI, configured tools and allowed project context.
You still have to build a good product. A reliable agent doesn’t appear because a model is reachable.
But the useful app shouldn’t have to become another isolated AI setup.
If you want to build one, the developer page has the SDK and the two ways to distribute.