I have spent a lot of the last two years messing with AI coding tools.

Some of that was just using them. A lot of it became trying to understand how LLMs actually behave inside coding workflows: context windows, local models, caching, retry loops, validation, and the architecture around all of it. I learned a lot from using different tools, but I also got tired of stitching the same pieces together myself.

That is where Aidesmith started.

The workflow I wanted

The first version started as a tool for myself. I wanted a native Mac workflow that kept the parts I liked from modern AI coding tools, but made the tradeoffs more explicit.

I ended up building around a few constraints I kept missing in other tools:

  • Cloud models for planning and the harder reasoning steps.
  • Local models for the repetitive edit loop.
  • Visible tasks, so the work is not hidden inside a chat transcript.
  • Validation built into the flow instead of treated as an afterthought.
  • Quota tracking, because token usage changes how you work.
  • Checkpoints, so experiments stay reversible.

Most of those are small product decisions, but they change how relaxed I feel when I let an agent work on a real repository.

Why local models matter

Cloud models are still useful for high-level planning and difficult reasoning. I do not want to pretend otherwise. But not every coding step needs the most expensive model in the loop.

A lot of implementation work is repetitive: inspect a file, make a small edit, run a check, respond to the result, repeat. When that loop can move to a local model, the workflow becomes cheaper and easier to experiment with. For my own work, Aidesmith cut cloud token usage by around 40-60%.

The surprise for me was how much this changed experimentation. When the cheaper loop is good enough, trying another model or prompt shape stops feeling like a budget decision.

The product idea

Aidesmith keeps the engineering process visible enough that you can steer it. The product gives attention to the small questions I kept checking manually:

  • What is the agent trying to do right now?
  • Which model is being used for this part of the workflow?
  • What changed?
  • Did the validation pass?
  • Can I get back to a known checkpoint?
  • How much quota did this path consume?

Those are the questions I kept asking while using AI coding tools seriously. Aidesmith is my attempt to make those questions part of the interface, not extra bookkeeping.

Where it is now

It is still a very early beta. I am using it on my own work, learning where the workflow feels good, and where it still gets in the way.

If you are interested in testing it, the project lives at aidesmith.app. I am especially interested in feedback from people who already use AI coding tools enough to notice the small workflow problems: context management, validation, cost, checkpoints, and model switching.