I built a desktop AI coding agent that runs 100% offline with Ollama

Oryn — Mon, 06 Jul 2026 08:10:21 +0000

I'm a solo developer with a full-time job and a family, and for the last several months I've been building an AI coding tool in the cracks of my day — early mornings, late nights, weekends. It's called Oryn, and I want to share why I built it and how it works, because the two decisions at its core might resonate with you.

The itch

Two things bugged me about the AI coding tools I was using:

My code went to someone else's cloud, and I paid a markup for the privilege. Every request routed through a vendor, wrapped in their pricing.
Everything was a subscription. Another monthly fee here, another there.

I wanted something that (a) let me use my own model — my own API key, or a model running entirely on my machine — and (b) that I could pay for once. So I built it.

What it does

Oryn is a desktop app for Windows. You open a project folder and describe what you want in plain English:

"Add a login page to my app with email + password and client-side validation."

It then runs an agent loop:

Reads the relevant files to understand your codebase
Plans the steps
Writes and edits files
Runs commands and tests
Verifies its own work

…pausing for your approval on each change if you want (manual mode), or running straight through (auto mode). Every step is visible — you can stop it, skip a step, or jump into the built-in editor and take over at any point.

Here's a real, unedited 2-minute run building a working app from a single sentence: https://youtu.be/D1utnumMAlo

The two bets

1. Bring your own model — or go fully local.
Plug in your own key (Claude / GPT / Gemini / Groq) and pay the provider directly, no markup from me. Or — the part I'm proudest of — run it 100% local with Ollama. No key, no cloud, nothing leaves your machine. For a lot of day-to-day work (refactors, boilerplate, tests) a local model is plenty, and it's private and free to run.

2. One-time price. Pay once. No subscription.

How it's built (for the curious)

Frontend: Electron + React + TypeScript, with a Monaco-based editor.
Backend: a bundled service that runs locally (bound to 127.0.0.1) and handles agent orchestration, file tools, terminal, git, and provider routing.
The agent loop is a plan → act → verify cycle with tool calls (file read/write, shell, git). Manual approval is a gate in front of every mutating tool call.
Local models run through Ollama; cloud models go directly from your machine to the provider using your own key.

What was hard (the honest part)

Making the agent stop at the right moment. Early versions would happily "fix" things I didn't ask for. Adding a visible plan + per-step approval solved both the trust problem and the runaway problem.
The local path is a balancing act. Local models are great for many tasks but still trail frontier cloud models on long multi-step reasoning. Letting you switch per task, instead of forcing one choice, was the answer.
Shipping a desktop app solo is a grind — packaging, signing, auto-updates, licensing. None of it is glamorous and all of it matters.

Try it / tell me I'm wrong

There's a 7-day free trial, no card: https://oryneye.com . It's a paid product (one-time, no subscription) — saying that up front so nobody feels baited.

I'd genuinely love feedback, especially:

How does the local Ollama path feel on your setup?
Where does the agent loop break for you?

As a solo dev, honest criticism is the most useful thing I can get. Thanks for reading. 🙏