The Angular CLI quietly shipped something few major frameworks offer: an official Model Context Protocol server. It first appeared in CLI 20.1 and has grown with every release since. Wire ng mcp into your AI assistant — Claude Desktop, Claude Code, Cursor, Copilot in VS Code — and it gains direct access to Angular's documentation, best-practices guide, project metadata, migration analyzers, and (in v22) your build, test and dev-server targets.

If you have been following the MCP story from the protocol introduction onward, this is a notable milestone. Frameworks have spent two years arguing about whether to integrate with LLMs. Angular skipped the argument and shipped a server.

This post sets honest expectations: what ng mcp actually exposes, what it doesn't, and where it earns its place in your workflow.

What ng mcp is — in one paragraph #

ng mcp is an MCP server that ships with @angular/cli. When started, it advertises a small set of MCP tools that an AI client can call. Those tools give the model authoritative access to Angular's knowledge (docs, best practices, a guided tutor) and project context (the apps and libraries in your angular.json). Since CLI 22.1 it also registers, by default, tools that run any angular.json target (build, test, lint, e2e) and manage ng serve dev servers; in 22.0 those were still opt-in experimental tools.

The Angular docs still describe it as experimental at the time of writing. Treat it as production-adjacent, not production-blessed.

What you get out of the box #

You don't normally start it by hand. Your AI client launches it as a child process with this command and talks to it over stdio:

npx -y @angular/cli mcp

Lesson 12.2 shows how to wire it into Claude, Cursor and VS Code. (Run ng mcp directly in a terminal and it just prints a sample client config and exits.) At the time of writing, the Angular docs list these tools:

The knowledge and workspace tools never change your project. The action tools are what let an assistant close the act → observe loop: propose a change, rebuild, read the errors, and fix them, all inside one session.

The tool set has changed across releases. Earlier versions shipped a find_examples tool and an experimental modernize tool, and up to CLI 22.0 the build, test and dev-server tools sat behind the --experimental-tool (-E) flag. From 22.1 the action tools are registered by default, and find_examples and modernize are gone. Always trust your client's live tool list over any article, this one included. Lesson 12.3 tours the current tools in detail.

Useful flags #

Two command-line options are worth knowing on day one:

CLI 22.2 also added --root <path> (repeatable) to limit which directories the server may read. The old --experimental-tool flag is now hidden and, with no experimental tools left, enables nothing.

Flags go at the end of the args array in your client config. A conservative everyday setup:

"args": ["-y", "@angular/cli", "mcp", "--read-only"]

And the full default tool set, for when you want the assistant to build and serve:

"args": ["-y", "@angular/cli", "mcp"]

Fair warning: run_target can run any target in your angular.json, including custom ones such as a deploy target. Start with --read-only, and review your targets before handing the assistant the full set.

Where ng mcp fits — and where it doesn't #

It is easy to assume that an official Angular MCP server means "AI can now see all my components, services, and routes." That is not what this is.

Here is the more accurate mental model:

If you want the AI to read your actual components and services, you pair ng mcp with a filesystem MCP server. They are complementary, not competing.

Why this matters even though it's small #

The tool set is intentionally narrow, and that is the interesting part.

Most framework-AI integrations to date have been scraped or hallucinated — the model knows about Angular because Angular wrote a lot of public docs, and the model read them at some point during training. That breaks the moment Angular ships a new API: the model is confidently wrong because its training data is frozen.

ng mcp flips that. When the model uses search_documentation, it is querying angular.dev right now, not last year's snapshot. When it uses get_best_practices, the conventions come from the guide the Angular team maintains. When it suggests a migration via onpush_zoneless_migration, the recommendations come from a tool the framework team wrote — not a guess based on a Stack Overflow answer from 2022.

The shape of the tool set is also a signal. Angular chose to expose:

• Knowledge access (docs, best practices, a guided tutor)
• Project introspection (project list)
• Migration analysis (zoneless plan)
• Build/test execution (opt-in until 22.0, default since 22.1)

This is the framework custodian's view of "AI assistance": keep the model anchored to authoritative sources, give it deterministic project queries, and let it close the act-observe loop on the developer's terms. It is the most conservative useful version of an MCP server a framework can ship, and that is a feature.

A typical conversation #

With ng mcp running and wired into Claude Code, a conversation might go like this:

The model is not guessing — it ran the analyzer tool, got a concrete next step, and produced a plan that maps directly to your code. That is what an MCP server adds over a raw chat box.

When to enable ng mcp — and when to skip it #

Enable it when:

• You're on Angular 20+ and the AI tooling you use already speaks MCP
• You're planning a zoneless or signal-forms migration and want a co-pilot for the audit
• You frequently ask AI assistants about Angular APIs and have been bitten by outdated answers
• Your team is standardizing how engineers use AI in the codebase

Skip it (for now) when:

• You're on an LTS branch of Angular and rarely touch the latest features
• Your AI assistant doesn't yet support MCP servers (most do in 2026, but a few enterprise IDE plugins don't)
• You want fully autonomous code edits — ng mcp is intentionally narrow; pair it with other MCP servers if that's the goal

What's next #

The next post in this series wires ng mcp into the four AI tools developers reach for most: Claude Code, Claude Desktop, Cursor, and GitHub Copilot in VS Code. After that we'll tour the tools in depth and compare ng mcp with generic MCP servers, extend it with a companion server, and finish with an end-to-end workflow.

The takeaway from this post is simpler: Angular shipped an official MCP server, it is narrow on purpose, and the narrowness is what makes it trustworthy. Treat it as a knowledge-and-analysis layer for your assistant rather than a remote control for your repo, and it will earn its ng keystrokes.

Try it yourself #

If you have an Angular 20+ project on your machine and Claude Code installed, the smallest possible try-it is one command:

# in your Angular project
claude mcp add angular-cli -- npx -y @angular/cli mcp --read-only

Then start claude in the same folder and ask:

That's the handshake working. Everything else in this series builds on it.