A dev workspace where the AI knows what you're doing editor, browser, terminal and agent all share context
The thing that's always bugged me about AI coding tools is that they're isolated. Your editor doesn't know what docs you have open. Your terminal AI doesn't know what file you're editing. You end up copy pasting context between five different windows just to ask a simple question.
Kit is my attempt to solve that. It's a single desktop window with a code editor (CodeMirror 6), a real Chromium browser, a terminal, a git panel, and an autonomous agent and they all share context. The AI in the terminal can read the page you have open. The agent uses your actual project files, not a blank slate.
The agent is the part I'm most curious to get feedback on. You give it a task in plain English, it plans and executes in a loop using file read/write, shell commands and project search. Every destructive action asks permission before running. You can drop a .kitrules or AGENT.md into any project and it reads your conventions at the start of every task.
There's also a visual pipeline builder called Stairs for chaining AI prompts, shell commands, HTTP requests and file ops together — handy for recurring workflows.
Works with OpenAI and Anthropic keys. MIT licensed, macOS and Linux.
GitHub: https://github.com/raiyanyahya/kit
Site: https://thekit.dev
This is exactly the direction things need to go. I’ve been running an autonomous setup locally (OpenHands + Debian 13 on an RTX 4080 Super) to build an open-core CMS, and the biggest bottleneck is always 'context fragmentation' between the browser and the terminal.
We’re currently experimenting with a 'Skills' folder architecture to solve this. Instead of hoping the agent 'understands' our Next.js 16 / Supabase stack, we have a dedicated directory that defines specific 'Capabilities'—like strict JSONB block generation or multi-currency logic.
Essentially, it’s a source-of-truth that our agents (and our internal Cortex AI package) must 'check out' before they start a task. It’s been the only way to keep 'Vibe Coding' from turning into 'Spaghetti Coding.
The shared context idea makes a lot of sense. That copy-pasting between tools is exactly where things break. The part that stands out is how much context the agent is actually using. There’s a balance between having enough to be useful and having so much that it starts making wrong assumptions. Also with the agent executing actions . .. even with confirmations, trust becomes a big factor. One bad action and people get cautious real fast.
How has it felt in real use so far, especially with longer or more complex tasks?
The product is pulling toward infra faster than the current name suggests.
“Kit” is clean, but too generic for what this is becoming.
Once the product moves past “AI dev workspace” and into shared execution context across editor, browser, terminal, and agent, the value shifts from utility to infrastructure.
That usually changes what the name needs to do.
Not “helpful dev tool”
More “core execution layer”
Because the thing that’s sticky here isn’t the UI.
It’s the shared context model underneath it.
That’s the part teams will actually adopt into workflow.
And that layer usually wants a name with more ownability than Kit.
Vroth.com would frame this much closer to infra.
Xevoa.com also fits if you want it to feel more productized.
Supercool
Thank you
The "Skills folder as source-of-truth" pattern from the first comment is exactly what I landed on too — but from a completely different direction.
I'm non-technical. No terminal, no Docker, no GitHub fluency. But the same problem: agent starts a task without knowing the conventions, the decisions already made, the patterns to follow.
My version is dead simple: a session-end MD file that captures what was built, what broke, what's next, what must not be forgotten. Agent reads it at the start of every new session.
No shared workspace. No tooling. Just a handoff file.
The interesting thing is we arrived at the same architecture — structured context that persists across sessions — from completely opposite starting points. Technical and non-technical founders have the same underlying problem. The solutions just look very different.
Launching today on PH.
https://www.producthunt.com/products/kit-6?utm_source=other&utm_medium=social