An agentic workstation
powered by
reverse engineering.

For code. Free. No account — bring your own key.

Windows x64 · macOS & Linux coming soon

$ codecraft-0.1.0-windows-x64.exe
Download ↓

Windows may flag the unsigned installer — click “More info” → “Run anyway”. Signing is on the roadmap.

  • Free — No subscription, no account
  • BYOK — Your provider, your key
  • Beta — 0.1 — research in public
  • Local-first — Keys sealed by your OS

bring your own key

Anthropic OpenAI Google Copilot OpenRouter DeepSeek Vercel Mistral xAI Groq Kimi Z.AI Qwen MiniMax Ollama LM Studio + 200 more via the models.dev catalog · or run local

No account. No markup. Your key, your provider, your bill.

what it is

  • Orchestrated agents
    Subagents run in the background while you keep working. Watch them, expand their progress, or kill them — one bar above the composer.
  • Permission modes
    Supervised, assisted, autonomous — you decide how much rope the agent gets on every run.
  • Review before you accept
    Changes land as reviewable diffs with per-file detail. Nothing merges until you say so — revert is one click.
  • Bring your own key
    Anthropic, OpenAI, Google, OpenRouter, Copilot. No account, no markup — your key, your provider, your bill.
  • Durable sessions & memory
    Restart mid-task; the session resumes. Context compacts into facts that persist across runs.
  • A real workbench
    Session tabs, document tabs, split editor, thinking and tool worklog cards — built around the agent, not beside it.

The story

One more thing.

CodeCraft began with a question you may have asked yourself: what exactly makes a frontier coding agent feel effortless?

To find out, we reverse-engineered the observable behavior of Devin's desktop client — built by Cognition AI, the team behind Devin and Windsurf. Sessions, permissions, context, subagents, memory. Studied as a system, not copied as code.

Then we re-implemented those behavior classes independently, on an MIT-licensed open-source runtime. No proprietary code, prompts, or assets. The result: measured improvements — like step-budget warnings going from 1/5 to 5/5 — the first of a series of evaluations now in progress.

This is a beta. The investigation into what makes an agentic IDE feel effortless continues — now with real users. Every session you run is signal; every bug you report gets fixed.

1/5 → 5/5 · step-budget warnings, measured

> evals running — warn_steps · ephemeral reminders · hooks · brain memory · subagent routing

> results published as they land — the team expects favorable numbers

> early adopters — your sessions are the dataset; tell us what works and what breaks: dev@codecraft.wtf

> public source release — in preparation

BYOK

Your key. Your model. Your machine.

CodeCraft ships no hosted provider and asks for no account. Add a key from any supported provider — it's sealed by your operating system's credential store and used only for your requests. You pay your provider directly, at their rates.

Anthropic · OpenAI · Google · OpenRouter · GitHub Copilot

Questions.

Is CodeCraft Devin? ▸

No. It's an independent re-implementation of behavior observed in Devin's client, built on open-source code. Not affiliated with Cognition AI.

Is it really free? ▸

Yes — no subscription, no account, no hosted middleman. You bring your own provider key and pay your provider directly for usage.

Is my API key safe? ▸

Keys are sealed by your operating system's credential store and only sent to the provider you configured.

Why release this? ▸

Part experiment, part wager. Proof of what AI-assisted reverse engineering can produce — and publishing the beta is the fastest way to learn what an agentic IDE should actually be. If you run it, you're already part of the research.

Is the source public? ▸

CodeCraft derives from an MIT-licensed open-source project — the license permits this release, and third-party notices ship with the app. A public repository is in preparation; publishing the source is part of the roadmap, not an afterthought.

What's next? ▸

Source publication, macOS and Linux builds, installer signing, voice input — and deeper research into orchestration, memory, and agent UX.