Introducing Codus — the code editor for AGI

The first time I watched four Claude agents land patches on four different repos at the same time — without colliding, without me babysitting — I knew the editor had to change.

Not the language. Not the keybindings. The substrate.

The thing IDEs got wrong about AI

Every IDE I've used in the last two years was built for one human, one cursor, one task. Then someone glued an AI assistant on the side. It can autocomplete. It can fix the file you're staring at. It cannot run the next four tickets while you sleep.

That's not the AI's fault. It's the editor's.

The editor assumes a single execution context. The chrome assumes one tab in focus. The shell assumes one prompt at a time. The whole operating model is built around the bottleneck — you — being the slow part.

In 2026, you are no longer the slow part.

Codus is built around the bottleneck moving

The bottleneck is now decision quality at the operator level: which feature to build next, which trade-off to take, which PR to merge. Everything below that — implementation, verification, code review prep — is parallelizable, and an autonomous agent can do most of it.

So we built an editor where the unit of work is the agent, not the file:

Four worktrees, mounted side by side, each with its own terminal, file system, and embedded browser.
One central planner that watches all four at once and proposes the next coherent batch of work.
A cloud bridge so a product manager can drop tasks straight into the right engineer's quadrant.
A status model — proposed → pending → running → done — that mirrors how teams actually move.

It's not a chat sidebar. It's an operations console for a small team of agents reporting to you.

What "the code editor for AGI" actually means

We are not predicting AGI is here. We're saying: when the agent gets meaningfully better — and it will, on a quarterly cadence — the editor it runs inside has to keep up. A 2010 IDE bottlenecks a 2026 model the way a fax machine would bottleneck Gmail.

Codus is the substrate that scales with the model. Smarter agent → more parallel work → less human typing → faster ship. Same operator, more leverage, every quarter.

Where we are today

macOS native — Apple Silicon optimized, code-signed, notarized. No Electron jank.
MCP-first — every tool is an MCP server, including the embedded browser, the file system, and the planner itself.
BYO keys — bring your own Claude / OpenAI / local model. We never proxy your tokens.
Free desktop, paid cloud — the desktop app is free forever. The cloud planner and team sync are $20 per developer per month.

Private beta is open on macOS. We admit new teams every Friday.

If your engineering velocity feels capped by how fast you personally can type, you are exactly who we built this for.