Your AI workflow builder shouldn't phone home. This one doesn't.
There's a quiet trade most of us have accepted with AI tools: convenience in exchange for our data leaving the building. Every prompt, every document, every workflow โ routed through someone else's servers, logged somewhere, subject to terms that change when they feel like it.
We wanted to see if we could have the good part without the trade. So we built C0fi.
C0fi is a visual workflow builder โ think n8n's drag-and-drop canvas โ that runs 100% on your own machine. No accounts. No telemetry. No cloud. It's a single HTML file talking to your local model through Ollama. Your flows, your files, your model. Nothing leaves your laptop unless you wire up a node that makes it.
Here's what C0fi does for you:
Privacy that isn't a settings toggle. It's architectural. There's no server to opt out of because there's no server. Your proprietary process, your client data, your half-baked ideas โ they stay local by default, not by permission.
No lock-in, no subscription. It's one file under an MIT license. Flows export as JSON you own. Finished flows export as standalone apps you can hand to a colleague. Nothing to cancel, nothing to migrate off of later.
But local doesn't have to mean primitive โ and this is the part we're proud of. In C0fi v6.8, the model doesn't just run inside the boxes. It builds the boxes. Describe a workflow in plain language โ "sample three answers, take a majority vote, then verify the winner against the actual rules" โ and the model wires up the nodes for you. There's a chat orchestrator that sees your whole canvas and wires it from your conversation. You can draw a flow by hand and ask it to extend what you drew.
And once a flow works, it can run itself โ schedule it every N minutes, daily, or weekly (Auto Brews), each run auto-saving its result. A workflow that quietly does its job while you do yours.
Nodes for the things you'd actually want: decisions, critic loops, web search, page reading, MCP tools, shared memory, live chat. Twenty-one demos ship in the box, including one where three independent samples reason over a puzzle, vote, and then a deterministic rule-checker verifies the winner โ because voting alone can launder a confident wrong answer, and the verifier checks the answer, not the model.
You can assign up to 5 different models for easy testing and bench marking.
None of it costs you a cloud bill or a data-sharing agreement.
If you've been curious about local AI but figured the tooling wasn't there yet โ it's worth ten minutes. You need Ollama, a mid-size instruct model, and the single HTML file. Load a demo, hit Brew, watch it run entirely on your own hardware.
Try it: Link in the Comments โ MIT licensed, no signup, nothing to lose but a little disk space.
What would you automate if you had and easy tool and everything ran on your machine?
Flies on a M4 Max 36gb Mackbook (quen3-coder:30b)
