I spent this weekend building with Switch for the Flip the Switch hackathon, and I walked away genuinely impressed. I want to share what I made and how it felt to work with it.
The idea behind Switch is simple once it clicks. Your AI agent usually sits alone in your terminal, where only you can reach it. Switch takes that same agent and drops it into the chat app your team already uses, so it shows up in a channel like any other teammate. Everyone reads from the same room, and the room keeps the history, the rules, and the decisions in one place.
I decided to build a Supabase connector. Plenty of teams run their backend on Supabase, and I wanted their agents to help with database problems without ever touching raw credentials.
What surprised me was how much Switch already handled for me. Adding a connector was mostly a matter of describing it. I set up the login flow, picked the read and write scopes, and wrote a small skill that tells the agent how to behave. I never had to build a client from scratch. Because connectors live in the core, the same Supabase connection works with Claude Code, Codex, or any other agent, and in any chat platform Switch supports.
Then I built the workflow I actually wanted. It is a small room where a person reports a bug, a triager agent reads the Supabase logs and advisors to find the cause, and a second agent proposes the exact SQL fix. The part I am most proud of is the approval gate. The fixer agent cannot run anything until a human approves the exact statement in the room. A thumbs up or a vague "looks good" is not enough. It has to be a clear yes to that specific SQL.
That one rule changed how the whole thing felt. The agents do the tedious work of digging through logs and writing the fix, but a person stays in control of anything that changes the database. It feels less like handing over the keys and more like working with a careful teammate who always checks before acting.
Not everything was perfect. Switch passes the tool list straight through, so it does not yet know which tools only read and which ones write. For now that guardrail lives in my skill instead of the platform. I would love to see read and write tiers built in, so an owner could grant read only access and have the system enforce it.
Even with that, I came away a happy builder. Switch took an idea that sounded complicated and made it feel reachable in a single weekend. If you have an agent sitting alone in your terminal, try giving it a seat in your team chat. You might be surprised how natural it starts to feel.
