The AL Performance MCP

At BC TechDays 2026, I did a workshop about AL Performance, where I covered 40+ exercises with attendees. I also had one thought, why not make MCP with knowledge of that workshop? So I did it, and it was premiered at BC TechDays 2026. Now it’s time to share with the rest of you.

If you’ve ever stared at a slow page in Business Central, opened SQL Profiler, and gone “why on earth is this fetching 100 columns for a lookup?” — this one’s for you.

So what is it, in one breath?

It’s an MCP server (Model Context Protocol — the thing that lets AI assistants call real tools) that plugs straight into GitHub Copilot, Claude Desktop, or Cursor. Once it’s wired up, you just talk to your AI the way you already do:

“Use al-performance to analyze this project.”

…and it reads your AL code, finds the performance sins, ranks them by how much they’ll hurt, hands you a phased action plan, and for a good chunk of them, just fixes them for you. No new window. No pasting code into some external analyzer. No leaving your editor. It all happens in the chat.

Why I built it

Here’s the thing about AL performance problems: they’re almost never exotic. It’s the same dozen mistakes, over and over, across every codebase I’ve ever reviewed. CalcFields() firing once per row. Count() <> 0 when a simple IsEmpty() would do. A record loaded without SetLoadFields, dragging every table extension column across the wire.

None of it is hard to fix once you see it. The problem is seeing it at scale, across hundreds of files, without spending your whole week doing code review by hand. That’s exactly the boring, repetitive, pattern-matching job that AI is genuinely good at. So, I taught it the patterns.

What it actually does

Let me just show you. Here’s the tool pointed at a real project, the same one my workshop attendees tore apart. One sentence in the chat, and this comes back:

67 files, 167 issues, sorted into 105 HIGH / 51 MEDIUM / 11 LOW and broken down by category — data transfer, locking, aggregation, memory, and so on. That’s the whole codebase x-rayed in a few seconds.

It tells you where to look first

167 issues sounds terrifying until the tool ranks them for you. Every file gets a weighted severity score (a HIGH counts for a lot more than a LOW), so you get a “worst offenders” list and know exactly which files are worth your afternoon:

Then it hands you a plan

This is my favourite part. Instead of dumping a wall of findings on you, it builds a phased action plan: Phase 1 is everything it can auto-fix for you (zero manual effort), Phase 2 is the high-severity stuff that needs a human, Phase 3 is the medium nice-to-haves. You always know what to do next.

And it explains itself

It’s not a black box shouting “BAD CODE” at you. Ask about any pattern and it’ll explain the root cause, show you the SQL the database actually runs in each case, and give you a before/after. It’s honestly a pretty great way to learn AL performance, not just patch it.

The numbers

38anti-patterns detected

12performance categories

11fixed automatically

Under the hood there’s a master orchestrator that fans the work out to specialist sub-agents — one for data transfer, one for FlowFields, one for locking, one for bulk operations, and so on — then aggregates everything into that single report. You can also drill into just one category if you want to focus.

And the 11 auto-fixable ones aren’t risky guesswork, they’re the safe, mechanical transforms. Things like turning Count() <> 0 into not IsEmpty(), Find('-') into FindSet(), adding a missing temporary keyword, or flipping ModifyAll(…, true) to false when you don’t need the triggers. It runs in dry-run mode by default, so you always get a preview before anything touches your files.

Getting it running

No clone, no pip install, nothing to build. It ships on npm and launches through npx, so you just point your AI host at it.

VS Code (GitHub Copilot agent mode) — drop this into .vscode/mcp.json:

{
  "servers": {
    "al-performance": {
      "type": "stdio",
      "command": "npx",
      "args": ["-y", "@bcility/al-performance-mcp"]
    }
  }
}

Claude Desktop — same idea in claude_desktop_config.json, under mcpServers. Cursor and other MCP hosts follow the identical npx -y @bcility/al-performance-mcp pattern.

Then just ask: “List all AL performance patterns.” If you see 38 of them, you’re live.

You’ll need Node.js 18+ (for the launcher) and Python 3.11+ on your machine, and any MCP-compatible assistant.

Go break it

This started as a workshop toy and turned into something I genuinely reach for on real projects. I’d love for you to try it on your own codebase and tell me what it catches, and what it misses. That feedback loop is how it gets sharper.

• npm: @bcility/al-performance-mcp

Point it at your worst codebase. See what falls out. And if you were at BC TechDays 2026, thanks for being the first to kick the tyres. 🚀