The Proof Pack.
One real Mendix™ app, authored end-to-end by AI through mxto, and every layer of evidence behind that, downloadable and inspectable. Not stronger claims; the artifacts themselves.
1. The reproducible loop, on Claudius
Claudius is a complete IT-asset / request tracker (25 entities · 68 microflows · 9 modules · MX 11.6.4) built by an AI agent driving mxto, no Studio Pro in the loop. The whole loop is reproducible from the artifacts below:
Every line above is verbatim captured output (Claudius round-trip, 2026-06-08). The companion "AI adds working logic live" sequence (author → mxbuild 0-CE → behaviour appears, DB-confirmed) is captured on the Comotion app. See the method.
2. Why your AI can read it: the same flow, three ways
This is the core claim, made concrete on one real Claudius flow (ACT_ApproveRequest,
all three files in the data pack). Your AI today gets the left column. With mxto it gets the right.
The raw model graph
"startNodeId": "2693b9bf-cf9d-…" "edges": [ { "originNodeId": "f6167db7-…", "destinationNodeId": "11aabc75-…", "caseValue": "true" }, … 8 more edges, each a pair of opaque IDs … ] # no order, no names
mxto typed intermediate form
flow: ACT_ApproveRequest
params: [Request]
nodes:
- guard: status in
(Submitted, InReview)
- change: Request.Status
= Approved
- commit: Request
- call: CreateAuditEntry
# typed + ordered
mxto read-optimized flow
flow ApproveRequest(Request): if not Request.Status in (Submitted, InReview): show_warning(…); return Request.Status = Approved Request.Reviewer = CurrentUser commit(Request) call(CreateAuditEntry) close_page()
Measured with tiktoken o200k_base;
resolution counts from the graph's own structure. Reproduce it + the full method:
04-comprehension-benchmark-method.txt.
The three files ship in the data pack: ApproveRequest-model-graph.json,
ApproveRequest.mir.txt, ApproveRequest.axon.
3. Download & inspect
The Claudius data pack
The Mendix project, the full model graph (our JSON), the readable flow + typed IR (shipped in the data pack), and mxto's own flow renders. Import it; rebuild it.
Tools & capability map
~140 MCP tools by phase, the read/write coverage map, and what we don't cover yet.
Mendix versions tested
Auditable: the exact versions (7.23 → 11.6), 100% round-trip, 4/4 build-gated, and what's not yet witnessed.
System requirements
A Mendix PAT + Project ID, a Mac/Linux/CI box sized to your model, and your own AI subscription.
What mxto doesn't do yet
The honest boundaries: what needs a human, and "structurally safe ≠ semantically correct."
Security & enterprise readiness
Local processing, the AI-agent data boundary, credentials, auditability, SOC 2 (not certified, by design).
Why not Mendix™ Maia?
How mxto and Mendix™ Maia fit together: the native IDE assistant alongside the external, headless delivery engine for estates, CI, and agents.
4. Prove it on your own estate
The strongest proof is your app, not ours. Send us a sample module (or run mxto locally) and we produce a read-only round-trip report: what extracted, the construct coverage, and the zero-loss diff, against your own model. No Studio Pro, nothing leaves your machine without your say.