The future belongs to directors.

The direction gap

Ask a model, working alone, what to build next, and the seam shows. Every choice it makes is synthetic — a weighted average of everything it has ever seen. One synthetic choice is fine. A hundred of them stacked is bland. A thousand is something stranger: a product where nothing is exactly wrong, but nothing quite fits. The parts are present, but the intent is missing.

What turns the same model into something exceptional is iteration with you. You push back. You recut. You say this is the part that matters, do more of it. The agent does the work; you give it direction. That loop is where judgment enters the system — where raw capability becomes purposeful, where output becomes intent, and where the work becomes worth shipping.

Making that loop fast, private, and natural is the entire point of Recursive.

The walled gardens are getting taller

If the next decade of software belongs to humans directing teams of agents, you'd think the industry would be racing to give people exactly that.

Instead, every model lab and every IDE is trying to make the cage prettier. Your prompts feed someone else's training run. Your agent's hard-won lessons about your codebase get locked into a vendor's proprietary memory store, where they belong to the vendor, not to you. Your data is scraped, anonymized, "aggregated" — words that mean we took it. If you want to switch providers, you start over. If a better model launches tomorrow, your workflow ages overnight.

This is the opposite of real creative work. Good direction needs continuity. It needs your old work, your old decisions, the small accumulated why of a thousand tiny choices. None of that should live on someone else's server.

What Recursive is

Recursive is the workspace I wanted and couldn't find.

It runs on your machine. Your code, your prompts, your sessions, your agents' memories — all of it stays where you put it. We send no telemetry. We store nothing on your behalf. The only thing that ever leaves your laptop is a request you, or an agent acting on your instruction, explicitly made.

It goes wherever you go. Your workspace lives on your hardware, but you don't have to. Kick off a run from your phone on the train. Check on your agents from a borrowed browser. Hand work off and pick it up from another room, another city, another timezone — without your code or context ever leaving the machine you trust.

It is model and provider agnostic. Claude today, GPT tomorrow, an open model the day after. Switch the engine without rebuilding the car. Your context, your plans, and your accumulated lessons travel with you.

It is fully extensible. Build your own UI on top of it. Define workflows we never imagined. Wire in a model provider that launched yesterday. Add tools, hooks, skills, plugins — every primitive is designed to be extended without waiting for us to ship a patch. If something is missing, you can build it. If something is in the way, you can replace it.

It is built agent-first, human-final. Tasks, plans, workflows, automations — every primitive is shaped so an autonomous agent can pick it up and run, while a human can step in at any moment to direct, approve, or course-correct. No more switching between five tools to ship a single feature. No more pretending an agent is a pair-programmer when what you really have is a team.

It is yours. A tool you run, not a service you rent. Your code, your prompts, your agents' memories stay on your hardware regardless of whether you ever pay us a cent. Supporters chip in once a year to fund the work directly — keeping Recursive independent of investors and ad networks, and keeping the roadmap accountable to the people using it instead of the people writing the next funding round.

What I want for software

I want a generation of builders who use this leverage to make stranger, more personal, more useful things — not the same fifteen SaaS apps with a chatbot stapled on. I want the small team that ships the thing the giant can't, because the giant can only build the median of what it already understands.

The models will keep getting better. The walled gardens will keep getting taller. The real question is whether the people doing the work — the ones with judgment, conviction, and a sense of what's worth making — get a tool that works for them, or get rented access to one that works for the platform.

The ask

Recursive is my answer. It's early. It's opinionated. It's independent, and I want to keep it that way.

If that's a future you want too, the most direct way to help is to become a Supporter. Supporters keep this project free of investors and ad networks. They get early builds and a seat in the room while the product is still small enough to shape.