Skip to content
arcade-agent

Claude Code skill (arcade-analyze)

The MCP server on the previous pages puts arcade-agent’s tools inside your agent’s loop. The arcade-analyze skill is the complementary packaging for Claude Code: a set of ready-made workflows that run the full pipeline — ingest → parse → recover → detect smells → compute metrics → visualize — and produce finished deliverables (an interactive HTML report, a drift report, an executive summary) instead of individual tool calls.

You don’t invoke tools; you ask in natural language. The skill triggers on requests like:

  • “analyze the architecture of /path/to/repo”
  • “compare PKG vs WCA on this project”
  • “what changed architecturally since v1.0?”
  • “explain the Clustering component”

and picks the right workflow. <source> can be a local directory or a git URL — the skill clones it for you.

What’s in it

Section titled “What’s in it”
  • Analyze — one codebase in, a self-contained HTML report out (component diagram, smells, metrics), auto-opened. An explorable app-style report — pan/zoom diagram, drill-down, dependency structure matrix, recommendations — is available as an alternative renderer. See a live report, generated from arcade-agent’s own codebase.
  • Compare algorithms — the same repo recovered with several algorithms (pkg, wca, acdc, arc, limbo), side by side.
  • Diff versions — architectural drift between two git refs: similarity, metric deltas, added/removed components, entity movements, new vs. resolved smells. The same idea the CI baseline runs on every pull request, pointed at any two tags you choose.
  • Query — summarize the architecture, explain one component, find where a concern lives, answer questions like “what’s the most coupled component?”
  • Architect deliverables — executive summary with a health score, findings, and recommended actions, plus further report-shaped workflows.
  • Architecture guardrail (arcade-guard) — the piece aimed at agents writing code, not just analyzing it. You describe the intended architecture in an architecture.spec.json (components by path glob, layers, allowed/forbidden dependencies), and deterministic checks keep work aligned to it: check gives a PASS/WARN/FAIL verdict with violations and fixes, propose says where a new piece should live before it’s written, preview answers “would this dependency be allowed?”. The same engine is also exposed as an MCP server (check_architecture, propose_placement, preview_impact, explain_violation, remediate), so any MCP agent can use the guardrail in-loop.

Install

Section titled “Install”

The skill is a thin wrapper over arcade-agent, so install arcade-agent first and point the skill at it:

Terminal window
export ARCADE_AGENT_HOME=/path/to/arcade-agent

Then, inside Claude Code, the repo is its own plugin marketplace:

/plugin marketplace add lemduc/arcade-analyze-skill
/plugin install arcade-analyze@arcade-tools

Prefer a bare skill? Clone and symlink it instead — the symlink name becomes the skill name:

Terminal window
git clone https://github.com/lemduc/arcade-analyze-skill.git
ln -s "$(pwd)/arcade-analyze-skill" ~/.claude/skills/arcade-analyze

Full workflow reference, options, and examples: github.com/lemduc/arcade-analyze-skill.

Skill or MCP server?

Section titled “Skill or MCP server?”

Use both — they answer different questions:

MCP server (arcade-mcp) Skill (arcade-analyze)
Shape 19 tools in the agent’s loop Packaged end-to-end workflows
Best at Mid-edit questions: context for a task, blast radius of a diff Deliverables: reports, drift diffs, summaries, guardrail checks
Output Structured data the agent consumes Artifacts a human reads (HTML, markdown)
Works with Any MCP client (setup) Claude Code