I love Claude Code. I built my own runner anyway.

I used Claude Code for a long time. I built a lot with it, and I liked it. This is not a breakup post. I still use Claude Code, but it stopped being my always-on daily driver.

Two things pulled me off it. The first is cost. I like it enough to say that without it being an insult, but an always-on metered budget adds up fast when an agent is in the loop all day. The second is that it is one vendor. I did not want to build a company on a runner I rented and could not change.

The clue

Somewhere in there I started using Ollama. On its own it was awkward. Then I pointed a runner at it, and it was nice to use. That was the clue. What I cared about was the runner, the loop and the tools and the guardrails, with whatever model I wanted underneath. A good runner pointed at a model I pay a flat rate for is a different cost curve than a metered frontier tool running all day.

So I built Merlin

Merlin is the runner I wanted. It is built on the things CorvidLabs already makes: fledge plugins are the tool surface, spec-sync specs are the contract. And it is provider-agnostic on purpose, like Cursor in the sense that you can use any AI. Today that is 31 providers across Anthropic, OpenAI-compatible, and Ollama. A tier picks the cheapest one that satisfies the job: most days I run merlin run --tier code, and it reaches for the cheapest provider that can actually do the work. The team runs on a yearlong Ollama subscription we already pay for, so day-to-day work is cheap and nobody has to think twice about using it. I bring in a frontier API when the task is worth it.

The honest split

I am not going to pretend I went cold turkey. Right now, contracting work still runs on Claude Code and Codex. When something needs to ship fast on a deadline, I reach for the tool that is already battle-tested, not an alpha I am still hardening. Claude Code is good at that, and I trust it there.

Day to day, the team runs Merlin on Ollama. It is consistent, it is ours, and it is cheap to run all day. Next month I will probably run Merlin with real OpenAI and Anthropic credits and make it the main driver. It already performs better on a paid model. The tier system means I choose, per task, instead of paying frontier prices for every keystroke.

Why it matters

I did not drop Claude Code as the always-on driver because it is bad. I dropped it because I wanted to own the runner and dial the spend per task, across any provider. And this is bigger than saving money. A runner that works with every model keeps getting better as the models do, and I would rather get the most out of all of them than tie the company to one.

Merlin is first-party to CorvidLabs, built on the tools we make. Not a second Anthropic tool. A third option that happens to be ours.