One AL Toolkit, Five Repositories

Note: This blog is a personal notebook. I write it to record what I work on, for my own reference, not to publish it on social media or to promote myself. If a post is still useful to you, I am glad, but that is not why it exists.

This post was published on August 15, 2026. Software versions, product features, links, and tool availability may have changed since then. Always check official documentation for the latest information. Some screenshots show small details. Zoom in on the page to see them clearly.

I write AL code for Business Central with four AI coding clients: the Copilot app, the Copilot CLI, VS Code Copilot Chat, and Claude Code. Every client wants the same AL review checklist in a different package format. A publisher knows the problem: one book ships as hardcover, paperback, e-book, and audiobook.

I did not want four copies of one checklist that drift apart. So I built five small public repositories. Four target one client each. One targets three clients from a single source, with the Agent Plugins 1.0 standard. All five are MIT licensed demonstration projects, not products.

The five repositories

Repository Client Delivery Install
mb-al-ai-toolkit-agent-plugin Copilot app, CLI, VS Code Agent Plugins 1.0 marketplace copilot plugin marketplace add, then copilot plugin install
mb-al-ai-toolkit-gh-copilot-app Copilot app Self-hosted marketplace Settings → Plugins → Add marketplace
mb-al-ai-toolkit-cli Copilot CLI Plugin from GitHub copilot plugin install mbaic/mb-al-ai-toolkit-cli
mb-al-ai-toolkit-cc Claude Code Plugin, own marketplace /plugin marketplace add, then /plugin install
mb-al-ai-toolkit VS Code with Copilot Chat .vsix extension code --install-extension

The first repository has its own post: Agent Plugins 1.0, one AL plugin for three Copilot clients. It serves three clients from one source and excludes a custom agent by design.

The other four carry near-identical content: a review skill, a refactor skill, a unit-test skill, naming rules, and a reference skill for RecordRef sorting. The Copilot app and CLI editions add a fast build agent. The VS Code extension is the simplest: no code, only markdown files and a package.json.

What changes per client

Two mechanisms explain almost every difference in the table.

Prompt files do not travel. VS Code Copilot Chat supports .prompt.md files. The Copilot CLI and Claude Code do not. A saved task becomes a skill with a clear description field, so the client can select it from a plain-language request.

A skill needs a call convention per client. In the Copilot app and CLI, a skill loads itself from your wording. In Claude Code, a skill also gets a namespaced slash command, for example /mb-al-ai-toolkit:al-review-code. The content behind it is the same file.

What I learned

The AL knowledge, the naming rules, and the checklist stayed almost identical across all five repositories. The packaging did the work, not the content. A skill is a markdown file with frontmatter in every client I tested, and the differences sit in the folder layout and the manifest fields.

Three habits saved me the most time:

  1. Write the description field for the client, not for a person. A vague description means a skill that never loads.
  2. Keep version numbers aligned across every manifest. A short validation script catches most mismatches before a release.
  3. Remember the cache. Every client copies a plugin at install time, so a push reaches nobody until they run the update command.

Closing note

If the Agent Plugins standard keeps its promise, three of these five repositories become unnecessary. I keep all five for now, to watch the standard hold up over the next months. Every repository is short. Read one, copy what you need, and fit the rules to how your team writes AL.

← Back to Agentic Coding