An AI coding agent that keeps your code on your machine.
A cloud planner sees file names and function signatures. A model on your own GPU writes the code, and your tests check every edit. Zuse lists every byte that leaves your machine.
Preview in early 2027.
How it works
Index on your machine
Zuse builds a symbol index of your repository locally. It is never uploaded.
Plan from an outline
The planner gets your request and an outline: file names, symbol names and signatures. Function bodies stay behind.
Edit on your machine
A model on your GPU or CPU makes each change, one function at a time, in a format that has to parse before it is written.
Check with your tests
Every edit has to pass your type checker and tests. If it fails, Zuse retries with the real error, then rolls the edit back.
Read the ledger
Every request that left your machine is listed, byte for byte. Undo a whole run with one command.
Measured, including the misses
Numbers from our development machine. Results on the two reference laptops come before the preview opens.
| What we measured | Result |
|---|---|
| Index a 100,000-line repository | 1.6 seconds |
| Re-index one changed file | 3 milliseconds |
| Planner request for a change to Zuse's own code | 14 KB, no function bodies |
| Local model finishes the task, given a correct plan, on an 8 GB GPU (Gemma 4 E4B) | 53 to 67% of 30 tasks |
| The same on a 4 GB GPU (Gemma 4 E2B) | 34% of 30 tasks |
The small local model is good at single-function fixes and weak at changes that span several functions. We publish that so you can decide where it fits.
What to expect on your hardware
| Your machine | What Zuse does locally |
|---|---|
| GPU with 8 GB or more, or a Mac with 16 GB of memory | Plans and makes multi-step edits |
| GPU with 4 GB, or CPU only | Answers questions, plans, and makes single-function fixes |
ISO 26262 for automotive software teams
Zuse is founded by an engineer with two decades in automotive software. We are opening design-partner pilots for automotive teams, from startups to mid-size suppliers, anywhere in the world, working under ISO 26262.
No coding agent can make code meet ISO 26262 on its own. What it can do is make sure every change it proposes goes through the verification your safety process already requires, and leave the evidence behind.
- In local mode, nothing leaves your machine, so code shared between OEMs and suppliers never moves.
- In the pilots, every edit is verified like hand-written code: your checker, unit tests, coverage for your ASIL and a named reviewer, so your team can classify Zuse at TCL1 under ISO 26262-8.
- Pilots include an evidence bundle for each change: requirement, plan, diff, checker and test results, coverage, reviewer and the ledger.
- Next: C and C++ editing with your coding-guideline checker (MISRA C, AUTOSAR C++14) as a gate on every edit.
- We do not claim certification; the tool classification and the safety case stay with your safety assessor.
We reply within two working days.
Early access and pilots
Tell us about your team. Automotive and regulated teams can ask about a design-partner pilot; everyone else joins the early-access list for the preview. We reply to pilot requests within two working days.
Questions
What exactly leaves my machine?
In local mode, nothing. With cloud planning on: your request, an outline of names and signatures, and short summaries of results. Never function bodies, unless you approve a specific snippet. The ledger in the editor lists every request.
Which models does it use?
The local model runs through Google's LiteRT-LM on your GPU or CPU, Gemma 4 today. Cloud planning goes through OpenRouter, routed only to endpoints with zero data retention.
What does it cost?
Local mode is free and needs no account. During the preview, cloud planning uses your own OpenRouter key, and each request shows what it cost.
Which editors and languages?
VS Code first, including VSCodium and Cursor. TypeScript, JavaScript, Python and Go today; C and C++ next.