Module 2: Install agent tools

In this module, you'll install the Moderne skills and the Moderne MCP server into your AI coding agent. You'll learn how to install everything at once, how to scope to a single agent like GitHub Copilot, and how to confirm the new capabilities are actually available.

Exercise 2-1: Install agent tools for every detected agent

Goals for this exercise

• Run mod config agent-tools install and understand what it changes
• Confirm the new skills appear in your agent
• Know how to roll back with uninstall

Key concepts

The Moderne CLI bundles two complementary kinds of agent tooling:

Component	What it provides
Skills	Procedural instructions that teach an agent how to create recipes, run them, and analyze impact
MCP server	A live process that exposes tools for semantic search, navigation, refactoring, and recipe execution

The install command auto-detects which coding agents are present on your machine and installs both components for each one. See Skills for AI coding agents and the MCP overview for full reference docs.

Steps

Step 1: Install agent tools for every detected agent

mod config agent-tools install

The CLI scans for installed coding agents (Claude Code, Cursor, GitHub Copilot, Windsurf, Sourcegraph Amp, OpenAI Codex), then installs skills and registers the MCP server for each one it finds. If no agents are detected, the CLI prints a message listing the supported agents and where it looked.

To remove everything later, run:

mod config agent-tools uninstall

Step 2: See where the skills landed

The skills are installed as plain markdown files. Their exact location depends on the agent. A few examples:

Agent	Install location
Claude Code	~/.claude/marketplaces/moderne/moderne/skills/<skill>/SKILL.md
Cursor	.cursor/rules/moderne-<skill>.mdc (per project)
GitHub Copilot	.github/instructions/moderne-<skill>.instructions.md (per project)
Windsurf	~/.codeium/windsurf/skills/<skill>/SKILL.md

For the full list see Supported agents.

note

Cursor and GitHub Copilot install skills per project (into .cursor/rules/ or .github/instructions/) rather than globally. If you want those skills available in another project, run the install command from that project's root too.

Step 3: Verify the skills are available

How you check depends on your agent. Restart the agent first if it was already running — most agents pick up new skills/MCP servers at startup.

Claude Code - Open Claude Code in your workshop directory and type /. You should see entries like /moderne:create-recipe, /moderne:run-recipe, /moderne:create-organization, and /moderne:analyze-impact.

Cursor / GitHub Copilot - Skills are loaded automatically as rules/instructions. Open the project in your editor and check that .cursor/rules/moderne-*.mdc or .github/instructions/moderne-*.instructions.md exist.

Windsurf, Amp, Codex - Skills are referenced by name in your prompt (for example, "Using the run-recipe skill, ..."). See Invoking skills for the agent-specific syntax.

Takeaways

• A single command (mod config agent-tools install) wires up every detected agent.
• Skills are just markdown. You can cat them to see exactly what your agent has been taught.
• The CLI updates these files when you upgrade. Re-run the install command after each CLI update.

---

Exercise 2-2: Scope the install to a single agent

Goals for this exercise

• Use a per-agent install command to scope to one agent
• Recognize when per-agent installation is the right choice

Steps

Step 1: Install for a single agent

If you only want agent tools for one specific agent, use a per-agent subcommand. For example:

mod config agent-tools copilot install

Other supported subcommands are claude, cursor, windsurf, amp, and codex. Each one installs both skills and the MCP server for that agent only. If the agent isn't detected on your system, the command prints a message and exits without making changes.

This is the right call when:

• You only use one agent and don't want files written for the others
• Cursor or Copilot is the agent you care about and you're inside the project where you want those instructions installed
• You're scripting the install in CI/CD for a specific agent

Step 2 (optional): Install only skills, not the MCP server

If you want the procedural guidance but not the live MCP server (for example, on a machine where you can't run a long-lived process), install just the skills:

mod config agent-tools skills install

This installs skills for every detected agent but leaves the MCP server unconfigured. You can pair this with Module 4 (Trigrep) to get fast search via the CLI without needing the MCP server.

Takeaways

• Per-agent install is useful for project-scoped agents like Cursor and Copilot, where the skills get written into the working directory.
• skills install is a lighter-weight option when you don't want the MCP server.

---

Exercise 2-3: Inspect a skill to understand what your agent now knows

Goals for this exercise

• Open one of the installed skills and read its content
• See for yourself how skills shape the agent's behavior

Steps

Step 1: Read a skill end-to-end

Open one of the skill files using the path that matches your agent. For Claude Code on macOS:

cat ~/.claude/marketplaces/moderne/moderne/skills/create-recipe/SKILL.md

For Cursor, look at .cursor/rules/moderne-create-recipe.mdc in the current project. For Copilot, look at .github/instructions/moderne-create-recipe.instructions.md.

The file is plain markdown describing the recipe development workflow: when to choose declarative YAML vs Refaster vs imperative recipes, how to scaffold a project, how to write tests with RewriteTest, and how to handle imports correctly. This is what the agent reads before responding.

The four skills you now have available are:

Skill	What it does
create-organization	Helps the agent assemble a curated set of repositories to test recipes against
create-recipe	Guides the agent through recipe type selection, scaffolding, and writing tests
run-recipe	Handles compiling, running, and diagnosing recipes against real repositories
analyze-impact	Turns recipe run data into reports and visualizations

See Available skills for the longer descriptions.

Step 2 (optional): Try a skill with a throwaway prompt

In Claude Code, invoke /moderne:create-recipe and give it a simple, throwaway request:

/moderne:create-recipe

I want to create an OpenRewrite recipe that renames the method getItems() to items() on com.example.ShoppingCart.

For other agents, see Invoking skills for the right invocation syntax.

Notice how the skill shapes the agent's approach: it picks a recipe type before writing code, scaffolds a project layout, and writes tests with the RewriteTest framework. Without the skill, you'd typically get a single YAML or Java file with no tests.

You don't need to keep the output. Stop the agent once you've seen it work.

Takeaways

• Skills are just procedural markdown your agent reads before answering. There's no magic.
• Reading a skill once is the fastest way to understand what behaviors your agent will fall back to.
• When the agent goes off the rails, you can compare its actions to the skill and see where it deviated.

Next up

In Module 3, you'll run Prethink against the workspace you set up in Module 1 and see what kind of pre-resolved context gets generated for your agent.