An AI-native IDE with a virtual software team inside it. You write the spec — a PM plans, a Coder builds, a Reviewer verifies. They keep going without you.
Closed beta — July 2026 · Public launch — August 2026
The problem
Copilot and Cursor are 1:1 assistants. You prompt, one model returns a diff, and you review it by hand. Every loop still runs through you.
Flat plans, credit meters and BYOK are all mixed together. You pay $20 and still can't say what a heavy month will actually cost.
They have to route through their servers — that is the business model. For a lot of teams, that fact alone disqualifies the tool.
The shift
It doesn't stop to ask. But nothing it does is irreversible.
How it works
The PM breaks your request into concrete steps and names the files it will touch — before a single line is written. You can step in and change the plan here.
The Coder runs a real tool loop: reading files, searching the codebase, editing several files atomically and running your test command — up to twelve rounds per attempt.
The Reviewer isn't handed a summary. It gets read-only tools and inspects the actual result, including compiler diagnostics. It can't approve code it never opened.
If the review passes, the work lands in your workspace in one atomic apply. If not, it goes back to the Coder with the findings attached — up to three times, without you.
Competition
Put two agents on the same role and they work the brief in parallel. A Judge then compares the results against criteria you wrote yourself — a plain markdown file on your disk, re-read on every single evaluation. Edit it and the next verdict changes. No black box.
Ghost branches
Every write lands in an isolated ghost overlay first, and reaches your real files only after the review passes — in one atomic apply, with a backup. The runner-up keeps its own branch, quietly finished in the background. A rejected attempt is one click away from being the one you ship.
Control
The orchestrator is a local daemon. Nothing about your code needs to pass through us — because we never built a server for it to pass through.
The model bill goes straight from you to the provider. What you pay us and what you pay for tokens stay two separate numbers — so a heavy month never turns into a surprise invoice.
Keys live in your OS keychain, never in browser storage. Point it at a local endpoint like Ollama and the entire multi-agent pipeline runs with no internet at all.
OpenAI, Anthropic, Google and Groq, plus any OpenAI-compatible endpoint. Give the Coder your strongest model and the Reviewer a cheaper one — they don't have to share a key.
Webview ↔ postMessage ↔ Extension Host ↔ WebSocket (localhost) ↔ Daemon
Status
This isn't a concept page. The architecture and the orchestration daemon are finished — the beta exists to prove them against real work.
Built on a fork of Visual Studio Code v1.127.0, so your extensions, keybindings and settings come with you.
Roadmap
Real-world validation, load testing and a security pass. Free for every beta tester.
Plans go live, with global and local payment options.
The low-cost platform licence for people who already own their API keys.
Shared credits, an admin dashboard, and rules and skills synced across a repo.
The closed beta opens in July 2026. Leave your email and we'll send you an invite when it does.
Used only to send the beta invite. No newsletter, no third parties.