Here's a 2.5 min demo if you want to skip ahead: https://www.youtube.com/watch?v=g_E6yhpN-zU
I've been using coding agents for a few years and have been frustrated by how quickly they jump to solutions. Even with plan mode and ask-user-question features, it's still a review dynamic. You describe a problem, they hand you a complete solution, you react. The solution looks plausible but may hide bugs or design flaws you can't catch, because you didn't think through the design yourself. Over time, your intuition for your own software degrades because you've stopped making meaningful decisions.
rubberduck flips this. Each session has three phases: design, implementation plan, and execution.
In the design phase, you describe what you want to build and rubberduck chats with you about how to solve it. I was trying to mimic the design conversations I used to have with colleagues at Microsoft and Stripe: sitting down, talking through options, poking holes in them, weighing trade-offs. rubberduck explores the design with you and investigates the codebase on your behalf, but you make every relevant decision. The output is a design document.
Once you're happy with that, the implementation plan phase takes over. The agent does a deep dive into the codebase to figure out how the design maps to actual code. This is similar to plan mode in Claude Code. It's mostly autonomous, since the meaningful decisions are already made.
Finally, execution. The agent runs in an isolated container using a dev environment image you provide. Each task is validated against your quality gates (tests, linting, formatting) before being committed. The output is a PR.
Pricing is $24/month with $5 of included credits (roughly one session, depending on complexity). Credits cover tokens (no markup) and execution compute (you specify the vCPUs/memory you need).
I've been bootstrapping for a few months and it's been fun to build. Would love for you to try it. Feedback very welcome.