# System requirements

> mxto is a single self-contained binary that runs on your machine. There is no mxto cloud service;
> **your model and credentials never leave infrastructure you control.** Here is what you need.

## 1. Mendix™ access
- A **Mendix Personal Access Token (PAT)** with model-repository read/write scope
  (`mx:modelrepository:repo:*`). Used only to pull and commit your model via Mendix Team Server.
- Your **App / Project ID** (the Team Server project the PAT authorizes).
- Provided via environment / `.env` / CI secret. Never written to mxto's output, never sent to a third party.

## 2. A machine to run it on
- **OS:** macOS, Linux, or a CI runner. No Windows requirement; no Studio Pro; no GUI.
- **The binary:** a single ~21 MB native executable (no Node/JVM runtime to install for the core
  read/author/query surface).

## 3. Memory & disk (scale to your model)
mxto reads the `.mpr` BSON directly, which is light. Rough guidance:

| Model size (example) | RAM (read / author) | Working disk |
|---|---|---|
| Small app: ~25 entities, ~90 flows (Claudius) | < 2 GB | tens of MB (extraction output) |
| Mid estate: ~350 entities, ~350 flows | ~2–4 GB | ~100 MB |
| Large estate: ~1,800 entities, ~4,000 flows (EFG) | ~4–8 GB (BSON-direct path) | a few hundred MB |

> **Caveat (honest):** the figures above are for mxto's own BSON-direct read path, which reads the
> `.mpr` directly and stays light. Older SDK-based extraction tooling can require a large JVM heap
> (~24 GB) on very large models. That heap is a property of that path, not of mxto's. If you only use
> mxto's native read path, the table above applies. (VERIFY exact figures against your own model with
> a dry-run.)

## 4. Your AI agent (bring your own)
mxto is the **toolchain your AI drives**, not an AI itself. You supply the agent and its subscription:
- **Anthropic**: Claude / Claude Code (the configuration we develop against), **or**
- a compatible agent/IDE (e.g. **Cursor**) or your own MCP-speaking client.

Because the agent is yours, **where your model text travels is your choice and your contract**: run a
local model and nothing leaves your network; run a hosted model and your model text follows that
provider's terms. mxto makes the choice yours instead of making it for you.

## 5. Optional: deploy & test locally
To take the loop all the way to a *running, tested* app on your own infrastructure:
- **Docker** (to build the MDA and run a Mendix runtime container), and
- the matching **Mendix runtime image** for your version (mxto orchestrates the build/run).

If you only read, author, and commit (letting Mendix Cloud build), this section is not required.