While building Vrykell, one question kept coming up: how much control should an AI agent actually have over your terminal?
Most AI coding tools operate in their own environment. But when I'm debugging a build failure on a remote server, I don't want to keep copying errors between my terminal and a chat window.
I want the agent to work with the terminal I'm already using, whether it's local or connected over SSH.
That's why Vrykell binds conversations to specific terminal sessions and preserves command output, working directories, git branches and exit codes.
I spent a week vibe coding a terminal. By the end, I had thousands of lines of code and a terminal I couldn't actually use.
I'm xtr-hub, and that's how Vrykell started. It's now an open-source terminal where AI agents can work directly on local or SSH sessions.
The first version was built with Electron. It wouldn't launch properly because require('electron') was returning a string instead of the API. I spent a day debugging it before moving to Tauri. Then I somehow ended up writing my own terminal renderer: ANSI parsing, screen buffers, Unicode handling, Vim resize bugs. AI kept generating fixes, and I kept running into new problems.
On August 7, I deleted 16,646 lines and started over.
I didn't stop using AI to write code. I stopped treating generated code as progress just because there was more of it.
The rebuild wasn't a straight line either. I redesigned the interface, reorganized the frontend more than once, changed how terminal sessions and workspaces operated, and eventually integrated ACP instead of trying to build every part of an AI agent myself. Some of the features I'd spent days building didn't survive those changes.
But through all those rewrites, the problem I wanted to solve stayed the same.
I'd SSH into a machine, run a command, and watch it fail. Then I'd open an AI chat, copy the error, explain which machine I was on, wait for a suggestion, and switch back to try the fix.
The next error meant doing it all over again.
I was tired of being the clipboard between my terminal and my AI.
So I built Vrykell around a simple idea: the agent should work where the problem actually happens.
A conversation is bound to a specific terminal, local or over SSH. Commands run on that target, and their working directory, git branch, output and exit code stay together as blocks. The agent gets the actual execution context instead of whatever I remembered to paste.
Files, an editor, git and SSH host management live in the same workspace. Vrykell also supports ACP (Agent Client Protocol), so Claude Code, Codex, Gemini CLI and OpenCode can work through the same interface, on the machine you choose rather than in a separate sandbox created by Vrykell.
On November 2, nearly three months after deleting the previous version, I finally shipped 0.1.0.
It's at 0.1.7 now. Still early, still plenty of rough edges. You bring your own model key. No account, no relay, no analytics. Open source under Apache-2.0, with installers for Windows, macOS and Linux.
https://github.com/Vrykell/Vrykell/releases/latest
If you give it a try, I'd love to hear how it works with your SSH setup, terminal applications or preferred AI agent. And especially where it breaks.
I've already rebuilt this thing more times than I expected. I'd like the next changes to come from people actually using it.