Hi all,

I have been using Claude Code with the official OutSystems MCP server on a real ODC project (a Mobile app published as a PWA, plus a library and a web app). The MCP server gives the agent the tools. What I kept missing were the working habits that stop those tools from producing quiet mistakes, so I packaged them as a free, MIT-licensed Claude Code plugin:

github.com/HoangAnhPTIT/odc-agent-toolkit

This is a personal community project. It is not affiliated with or supported by OutSystems.

Why: a fast, reliable, standard way to work on ODC with an AI agent

Without a shared workflow, every developer and every agent session works with ODC differently, and the same small mistakes come back. The goal of the toolkit is one process that is:

• Fast: the agent knows up front what MCP can and cannot do for each asset type (web app, library, Mobile app, workflow), so it picks the right path instead of trying and failing.
• Reliable: every change follows the same loop: read the current model, design the change, edit, confirm with you, publish, then verify what ODC actually stored against what was asked. A successful turn or publish is treated as a signal, not as proof.
• Standard: the same rules, the same templates and the same checks in every project and every session, written down in one place you can review, version and improve.
• Complete: it also covers what MCP cannot do. Mobile apps cannot be edited over MCP, so their changes become a step-by-step ODC Studio guide with click-to-copy values and an automatic check; library release and dependency updates become explicit steps instead of being forgotten.

How it is kept honest

Every rule in the skills cites the public ODC docs source or an MCP tool description; anything that comes only from experience is marked "observed". The scripts have unit tests, and the plugin is improved by running it on real tasks and fixing whatever slowed the agent down or let a mistake through.

Install

You need Claude Code with plugin support, the OutSystems MCP plugin connected to your tenant, and Python 3.9+. In a terminal:

claude plugin marketplace add OutSystems/outsystems-mcp

claude plugin install outsystems@outsystems --config tenant_hostname=.outsystems.dev

claude plugin marketplace add HoangAnhPTIT/odc-agent-toolkit

claude plugin install odc-agent-toolkit@odc-agent-toolkit

(Inside Claude Code the same works as /plugin marketplace add … and /plugin install ….)

Restart Claude Code, then ask in plain language, for example "Read the Orders screen of  and explain how it works", "Change this action in , publish it and check the published version", or "Write a Studio guide for this change to the mobile app". The repository includes a filled example guide for a fictional app.

Limitations

• context_graph shows published revisions only and cannot resolve some references (client variables, static entity records, elements of other modules).
• The graph is the model, not the runtime: behaviour still has to be tested.
• Tool names and behaviour come from the OutSystems MCP server and may change between releases.
• Raw graphs contain your app's model: keep them out of public repositories.

Feedback welcome

Issues and pull requests are welcome on GitHub. New rules need evidence (a docs page, a tool description, or a reproducible observation). I am especially interested in other Mentor or MCP gotchas you have hit, and in how others handle Mobile apps.

Thanks!