Agentic coding, with the loop closed

What the agent does itself

1. The compiler is part of the loop

   A compile takes 2 milliseconds, so being wrong fifty times costs a tenth of a second. Every error names the file, the line and what was expected, clearly enough for the agent to fix it alone. Programs are written in Wantzel: strict, one way per concept, and a compiler without dependencies. How the compiler works and why it is so strict.

   Developer mode

   

2. Many agents, one board

   Start as many agents as you have tickets, each with its own job, all in one overview. Hand a ticket to an agent and it writes back in the work log what it did. A stuck group stops in one go.

   Thirty-two agents at once

   

3. Your own AI at the controls

   The studio is an MCP server. What you do in the window, Claude or any other MCP client can do too: open a ticket, start an agent, read what came back. The agents themselves run on the models you choose, each on your own account.

The questions people ask

• Which agents run?Claude, Codex, Copilot, Gemini and open models such as DeepSeek and Kimi, each on your own account. If one hits a limit, the run goes on with another.
• What do I host?Nothing. The studio runs natively on its own server in Germany (EU), with publishing, your own domain and the certificate arranged.
• Can I still write code myself?Yes: the editor, the debugger and the build are all yours. The agents work in the same files and the same tickets.
• What does it cost?EUR 200 a month for your whole organisation, with no mark-up on tokens.