# mxto tools & capability map

> What your AI agent can actually *do* to a Mendix™ app through mxto: ~140 MCP tools plus a CLI,
> grouped by what they're for. Then the capability map: what mxto reads, what it writes, and (just as
> important) **what it does not cover yet.** A single ~21 MB binary; Mac / Linux / CI; no Studio Pro.

## The tool surface (by phase)

| Phase | What the tools do | Examples |
|---|---|---|
| **Onboard** | Orient an agent and check environment health before work | start-here, readiness, environment doctor, version report |
| **Extract** | Read a `.mpr` directly into a typed, queryable model (the read foundation everything else builds on) | extract model, warm corpus |
| **Query** | Ask precise questions of the model: entities, attributes, associations, microflows, pages, navigation, security, call graphs, cycles, duplicates, expressions, type catalog, generated SQL | query entity / flow / module / page / graph-stats / SQL |
| **Query lenses** | Higher-level semantic projections that compose raw queries into a usable view | flow lens, entity lens, impact lens, module lens, security lens |
| **Comprehend** | The read-optimized projection an LLM reasons over cleanly (in order, by name), plus per-flow intent briefs | Axon projection, flow-context briefs |
| **Certify** | Prove a change is safe: round-trip fidelity, construct coverage, dead-code, lint, quality score, security/RBAC audit, and deep semantic-equivalence / invariant verification | round-trip, coverage, dead-code, lint, security audit, equivalence checks |
| **Transform** | Emit derived artifacts from the model: SQL (DDL + SELECT) and model-to-specification compilation/translation | SQL DDL, SQL SELECT, spec compile |
| **Write (model)** | Author every construct type back into the model and commit: domain, microflows, pages, security, resources | author domain / microflow / presentation / security / resource; apply; commit |
| **Write (source)** | Author source-level details: entity documentation, indexes, attribute defaults | entity doc, entity index, attribute default |
| **Debug** | Live debugging of a running app: breakpoints (including by name), step in/over/out, inspect variables and objects | set/clear breakpoint, step, inspect object |
| **Runtime** | Drive a running Mendix runtime: retrieve / create / commit / delete objects, execute microflows, read metadata & enums | runtime retrieve / create / commit / execute |
| **Admin (server)** | Operate a live server: status, database health/queries, log search & stats, scheduled events, slow queries, lock diagnosis | server status, db query, log search, slow queries |

*Exact per-phase counts and the full tool list are reported by the binary's own `onboard start-here`
on install. This map is the high-level shape, not the registry.*

## Capability map: read vs write

| Direction | Coverage | What it means |
|---|---|---|
| **Read** (Mendix → typed model) | **100%** across 8 production apps, **0 transform failures** | Every entity, attribute, association, microflow, page, security rule and integration becomes typed records you can query. |
| **Write** (author → commit → re-extract) | **70 / 70 construct types certified** (0 non-automatable) | Entities, attributes (9 types), associations, enums, generalizations, validation rules, event handlers, indexes, module/user roles, project security, access rules, constants, scheduled events, datasets, Java/JS actions, published REST, modules, pages, snippets, navigation, layout grids, authored and round-trip-verified. |

"Certified" = authored, read back, and schema-compared against a Studio-Pro-authored golden
(wire-level fidelity); a representative subset is also proven through a real `mxbuild` to a running app.

## What mxto does NOT cover yet (stated plainly)
- **Fully offline / air-gapped page-UI authoring** is on the roadmap. Page UI is authored today through
  the standard write path; the rest of the offline write surface (domain, security, navigation,
  microflow bodies) is already in place.
- **Deep custom content in complex pluggable widgets** (e.g. nested content inside an advanced data
  grid) still has edges that need the Studio Pro UI.
- **Java action source bodies**: the action signature is authored; the Java body is hand-written.
- **Published SOAP / consumed OData** read is shallow today, not yet deep-certified.
- **Some integration sub-properties** (e.g. certain REST-call HTTP configuration fields) have a
  documented manual fallback.
- **Not every Mendix minor version** is witnessed. See `01-mendix-versions-tested.md` for the exact list.

> None of the above is a *missing construct type*: read coverage is 100% and all 70 write construct
> types are certified. These are deep sub-property and offline-path edges, tracked openly so the map
> survives technical due diligence. Semantic correctness (whether a structurally-valid change *means*
> what you intended) is always a human/verification decision (see `boundaries`).