Skip to main content

Getting Started

Install the AI-Pack plugin into Claude Code, set up the knowledge-graph server, and run your first multi-agent workflow.

Prerequisites​

  • Claude Code installed and working
  • Git (to clone the repo and initialize the mcp submodule)
  • Python 3 (used by the kg installer)

Installation​

1. Clone the repo​

git clone https://github.com/Cortexa-LLC/ai-pack.git
cd ai-pack

2. Install the plugin​

make install-plugin

This registers the local marketplace with Claude Code and installs ai-pack@ai-pack. It is a one-time step; the plugin then works in any project you open with Claude Code.

3. Set up the kg knowledge-graph server​

The plugin's .mcp.json launches a kg MCP server that gives agents persistent memory. kg is a standalone binary provided by the mcp git submodule:

git submodule update --init mcp
python3 mcp/install.py --mcp kg

See Knowledge Graph for what it does and how agents use it.

4. Restart Claude Code​

Close and reopen Claude Code so the plugin and MCP server load.

Verify the installation​

Agents. In a Claude Code session, the six subagents should be available to the Agent tool as ai-pack:* subagent types: ai-pack:architect, ai-pack:engineer, ai-pack:inspector, ai-pack:pr-shepherd, ai-pack:reviewer, ai-pack:spelunker. Ask Claude to "list the available subagent types" if you want to confirm.

Knowledge graph. From a terminal:

claude mcp list

The output should include the kg server. You can also run scripts/verify-kg.sh from the ai-pack repo.

Skills. The three skills appear as /ai-pack:orchestrate, /ai-pack:pre-push, and /ai-pack:shepherd-pr, and also trigger automatically on matching requests.

First use​

Open Claude Code in any project and describe work in plain language — the skills and agents pick it up:

Drive a PR to merge-ready:

shepherd PR #42 to merge-ready

This triggers the shepherd-pr skill, which spawns the pr-shepherd agent to watch CI, fix failures, and address reviewer threads until the PR is green and approved.

Multi-step engineering work:

investigate why the login flow intermittently fails, then fix it

This triggers the orchestrate skill, which decomposes the request — an inspector agent finds the root cause, then an engineer agent implements the fix with tests.

Check your commits before pushing:

review my commits before I push

This triggers the pre-push skill: a reviewer agent examines your local diff, and if issues are found, an engineer agent fixes them and amends the commit, looping until the review passes.

Updating​

After pulling changes to the ai-pack repo:

make update-plugin

Next steps​

  • Agents — what each subagent does and when to use it
  • Skills — the three workflow skills in detail
  • Knowledge Graph — persistent project memory
  • Task Packets — optional convention for multi-session briefs

Support​