Mendix 11 turns Maia from a scattered set of AI helpers into a single, context-aware co-developer that plans, builds, styles, and refactors real applications – and it pairs that with a production-grade GenAI stack for putting AI features into the apps you ship. In practice this means you can go from an idea to a running, editable Mendix app largely by describing what you want, then build and deploy your own AI agents on top of it.
Maia Make: one chat that builds your app
A unified AI development chat in Studio Pro. Maia Make consolidates what used to be separate assistants – domain model, pages, chat, and explanations – into one conversational interface, enabled by default. From a single chat you can describe a data structure and get entities and associations, generate pages, produce logic, and ask for plain-language explanations of how your app works. Maia also thinks before responding, giving real-time visibility into its reasoning, and can enter a plan mode where it asks open-ended, single-choice, or multiple-choice clarifying questions before it generates anything.
Broad generation across the model. Maia Make generates and edits a wide swath of the Mendix model from natural language, including:
- Domain models – entities, attributes, associations, and cross-module references, plus non-persistent entities.
- Microflows – activities, XPath constraints, expressions, and CRUD logic – and it can generate matching unit test and validation microflows in the same conversation.
- Workflows, including Multi-user task elements, AI Tasks, Event Subprocesses, and Message Events.
- Pages built from scratch or modified, working with pluggable widgets and their design properties, plus app-level Navigation and Menu documents.
- Whole modules and folder structures, organized to your existing layout or to Mendix best practices.
- Enumerations, constants, and OQL – including creating and refining view entities where Maia writes the OQL for you.
- Data transformations (JSLT), and read-only support for Java and JavaScript actions so Maia can use and explain them (and create JavaScript actions, parameters, and their associated files).
Refactoring, error-fixing, and safe iteration. Maia does more than create – it can remove document elements and rename entities, attributes, and microflow parameters, and set annotation texts in the logic editors. Generated documents are automatically checked and their errors fixed, cutting the manual cleanup that AI output usually requires. When you want to walk changes back, undo works per document so you can revert a single generated document without losing the rest of the session, while Undo All reverts an entire session at once.
Rich context in and out of the chat. Maia grounds its work in far more than the prompt text:
- Paste screenshots, wireframes, or snippets directly into the chat, and attach Markdown files such as architecture docs and specifications (up to 50 KB each) as context.
- Fetch and interpret static web content from shared URLs – external documentation, JSON specifications, and public documents – with a preferences toggle to disable web fetch that now takes effect instantly.
- Read the System module and installed marketplace modules for context, and cite sources from the Mendix Support Knowledge Base in its answers.
- Search the Marketplace and download or install modules without leaving the conversation.
Custom instructions and reusable skills. You can steer Maia with project- and module-level custom instructions (Maia Agent Instructions, via AGENTS.md), applied automatically to conversations or when Maia works with a given module. You can also define your own Agent Skills as SKILL.md files – with optional Markdown reference files for domain knowledge – and these skills now ship and export with their module, making them reusable across applications. A Skills pane lists every custom skill with its status and module ownership, filters to skills with errors, and lets you sync updates immediately, while Maia indicates which skill it is using.
AI-powered styling and design
Maia designs the look, not just the structure. Maia accepts PDF brand guidelines, Figma design tokens, or plain descriptions and translates them into working CSS, modifying theme variables and design properties for colors, typography, spacing, and border radius – with built-in WCAG contrast checks. It can create custom styling and manage design properties as reusable styling variants, and it assists with Atlas 4 migration from SASS to CSS variables.
Navigation and menus by description. Maia creates and manages navigation profiles and menus, including device-specific navigation for phone, tablet, and responsive web, and Progressive Web App settings. It manages page parameters intelligently, prevents duplicate profiles, and can build menu items from the pages in your app.
Maia Plan: from idea to backlog
An AI planning engine in the Mendix Portal. Maia Plan turns unstructured inputs – notes, briefs, discussions – into structured project scope with goals, success criteria, personas, requirements, and boundaries, then generates Epics, structured User Stories, acceptance criteria, and technical recommendations. You can upload up to three files to enrich a plan and use collaborative Scope Refinement to close gaps and improve completeness. It targets the roughly quarter of delivery effort teams spend clarifying requirements.
Straight into building. Plans connect directly to implementation and to your tooling:
- Hand a plan to Studio Pro with a click, sending selected user stories to Maia Make; in Studio Pro you can also manually link and realize existing user stories throughout development.
- Push plans straight into a Jira backlog as epics and user stories, configured once, and switch between Jira and the Mendix Planning Board in project settings.
- Get automatic Marketplace component recommendations based on generated user stories, respecting Company-Approved standards when enabled.
Building AI agents in Studio Pro
Agents are now a first-class part of the app. With Agents Kit 2.0 and the Agent Editor as the preferred design-time approach, building an agent replaces manual microflow assembly and becomes part of the app model with version control and deployment. The Agent Editor separates a Build view for configuring an agent’s type, model, prompts, and tools from a Playground view for testing agents with variable substitution. You can build both task and chat agents, define system and user prompts parameterized with context objects, wire in tools built from microflows and consumed MCP services, integrate knowledge bases, and call an agent from business logic with a Call Agent microflow action. Maia can even refine an agent’s system prompt on request.
Control and observability. Human-in-the-loop approval steps let you gate tool execution until a user approves, with visibility configurable per tool and per knowledge base, and agent traces give full visibility into tool calls. Fast reload applies agent changes without a full restart, tightening the iteration loop.
- Agents Kit ships ready-to-use agent templates, an Agent Commons module for runtime prompt iteration, tool linking, and knowledge base integration, and reusable Conversational UI snippets.
- Reference multi-agent patterns – prompt chaining, prompt chaining with a gatekeeper, evaluator-optimizer, and routing agent – are available as showcase implementations, and function calling supports agent-controlled computer operations like clicking, typing, and navigation.
Model Context Protocol: Mendix as client and server
Mendix speaks MCP in both directions. Your apps can act as MCP servers, exposing Mendix-built tools (including microflow-based tools) to external AI clients, with custom microflow-based authentication and automated setup with schema generation. The MCP Client lets apps and agents call tools exposed by external MCP servers, spanning GenAI providers such as Amazon Bedrock, (Azure) OpenAI, Mistral, and Google Gemini plus bring-your-own-LLM endpoints. Both modules use Streamable HTTP transport, the current standard (the older HTTP+SSE transport is deprecated).
Studio Pro itself is an MCP endpoint. Studio Pro functions as an MCP server that exposes Maia’s generation tools – including page-editing tools – so external clients like Claude Code or Copilot can create microflows, domain models, and pages via your own AI agents, with changes reflected back in Studio Pro. Maia’s own MCP Client can connect to servers such as Playwright and Figma to ground its decisions in external context.
- MCP connections support OAuth (with a dedicated callback server) and Mendix Identity authentication, including automatic platform access tokens for Mendix-hosted servers.
- Mendix Cloud can be managed with AI through MCP integration, letting agents interact with cloud operations.
The GenAI stack for your apps
A managed, production-ready GenAI foundation. Mendix Cloud GenAI provides text generation, embeddings, and managed vector databases for retrieval-augmented generation and semantic search. A single resource now exposes multiple Claude model families – Haiku, Sonnet, and Opus – behind one API key, with new model versions added automatically; you can set default models per resource and select models dynamically. Alongside Anthropic Claude Sonnet, Cohere Embed (with 100-plus language support) and a range of other models are available, and you can bring your own via connectors.
Provision and scale it yourself. Admins can provision GenAI resources from Control Center using Cloud Tokens in flexible amounts (no fixed capacity tiers), self-service scale text and embedding plans (upgrades immediate, downgrades at month-end), and add external collaborators (who can manage most resource operations, but not team membership or environment settings). Consumption moves to a flexible capacity model of GenAI Units, where one Cloud Token equals 100 GenAI Units per month.
Connectors, knowledge bases, and observability. Platform-supported connectors cover Mendix Cloud GenAI, OpenAI, Amazon Bedrock, and Mistral AI, with capabilities including document chat, function calling with multiple primitive input parameters, and controllable tool choice. Knowledge base integration is vendor-agnostic through a common entity, so you can switch between providers – Mendix Cloud, Amazon Bedrock, Azure AI Search, or a custom PgVector implementation – without changing core logic, and a single bot can draw on several knowledge bases at once. LLM trace observability in the Conversational UI module shows API calls, tool calls, and knowledge retrievals with a daily invocation diagram, error counts, and drill-down traces – visualized as span breakdowns and as a Gantt chart that turns an AI’s chain-of-thought into an intuitive timeline.
- A public-beta setting lets Maia itself run on a custom LLM backend – OpenAI-compatible providers or AWS Bedrock – for cost control, data residency, and model choice, while the Mendix Platform provider stays recommended.
- Starter and showcase apps – a Blank GenAI template, the Agent Builder Starter App, the GenAI Showcase (now built on Studio Pro agents), and an RFP Assistant Starter App – give ready foundations on Mendix Cloud GenAI.
AI beyond development
Maia across the platform. AI assistance extends past Studio Pro into the surrounding delivery lifecycle:
- App Insights uses a Maia Feedback Processor to categorize and group incoming feedback and turn it into clear, actionable, backlog-ready user stories, and can link related feedback to existing stories to avoid duplication.
- Compass offers AI-guided vision and goal setting, asking clarifying questions and delivering a personalized low-code strategy vision statement.
- A machine-learning showcase demonstrates integrating predictions – such as employee-attrition scoring – via REST services and AI Hub endpoints without deep data-science expertise.
- Pluggable widgets can declare a prompt element describing their purpose and properties so Maia gives smarter, widget-aware assistance.
- A Capture Maia Agent Session support tool records a session for troubleshooting when raising support cases.