Agent Plugins 1.0: One AL Plugin for Three Copilot Clients

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 12, 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.

GitHub published Agent Plugins 1.0 on August 6, 2026. VS Code, the Copilot CLI, and the Copilot app all read it by August 12. The standard packages a skill and an MCP server once for all three, so one AL review checklist serves three clients from one repository.

Problem

Each client used its own file layout, so a team kept three copies of one checklist, and the copies drifted apart. It was three phones with three chargers. Agent Plugins 1.0 is the one plug that fits all three.

What moves, and what stays behind

Two things are portable: a skill and an MCP server. Write each one once, and every supporting client reads it.

A custom agent is not portable. The specification places custom agents outside the 1.0 format. A client-specific folder can hold one, with no guarantee across clients or versions.

I shipped a custom agent first, in a folder name I invented myself. Two clients loaded the skill and the MCP server and silently skipped the agent. No error appeared. I found the mistake only by comparing the package against an older one that worked.

GitHub documents a flat agents/ directory at the plugin root. That fixed the load failure, but not the missing guarantee, so I dropped the custom agent. The skill already carries the review checklist and the output format in every client.

Architecture

The plugin lives in one repository, mb-al-ai-toolkit-agent-plugin, which is both the marketplace and the source.

Component Type Function
al-code-review Skill AL review checklist: data access, object structure, events, permissions, tests, upgrade code
microsoft-learn MCP server Public Microsoft Learn endpoint. Read-only.

The mcp.json file needs an exact $schema value and the transport name streamable-http, not http. I shipped http first. A strict client silently drops an unknown transport and still loads the skill, so a small validator now checks the manifest before every release.

Install it

Each client keeps its own copy of an installed plugin, so each one needs its own install step.

Client Command or setting
Copilot CLI copilot plugin marketplace add mbaic/mb-al-ai-toolkit-agent-plugin, then copilot plugin install bc-al-toolkit@mb-al-ai-toolkit-agent-plugin
Copilot app Settings → Plugins → Add marketplace → mbaic/mb-al-ai-toolkit-agent-plugin, then install bc-al-toolkit
VS Code User settings.json: "chat.plugins.marketplaces": ["mbaic/mb-al-ai-toolkit-agent-plugin"]. This is a user setting. A workspace or dev container setting does not work here.

You call no command to use the skill. Ask for an AL review in plain words, and the description field in SKILL.md decides how well the client matches your request.

Risks and limits

  • Read-only is not the same as low risk. The plugin sends your proprietary AL source into the model context, so check your data policy first.
  • A push to the repository reaches nobody until each client updates its own copy.
  • The plugin does not compile or test AL. The AL-Go build and a human review still decide the outcome.

Closing note

Two things carry the value: the checklist and the read-only documentation lookup. Both now move across clients. One plugin, one skill, and one MCP server are enough. A small scope keeps a hobby project a hobby.

← Back to Agentic Coding