AI agent plugin & skill¶
ArchMAP ships as a ready-made plugin for Claude Code and as a portable agent
skill for Cursor, Codex, OpenCode, and other agents that read SKILL.md.
Installing it replaces the manual MCP and hook setup described in
archmap mcp and archmap memory.
Prerequisite¶
The plugin calls the archmap CLI, so install it first:
Claude Code¶
The plugin provides:
| Component | What it does |
|---|---|
MCP server archmap |
Runs archmap mcp . and exposes all fifteen ArchMAP tools (architecture, context, quality, findings, reverse engineering, audit) |
SessionStart hook |
Injects archmap memory --print into context at session start. Skips silently if archmap isn't installed, and is bounded to 60 s |
archmap skill |
Tells the agent when to call which tool: orient with get_project_memory and get_relevant_context, check get_analysis_quality, run impact_analysis before editing, run_checks + diff_architecture before finishing; plus a reverse-engineering workflow (inspect_path, extract_strings, hexdump) and audit_project |
To also refresh .archmap/memory.md after every edit, add the
PostToolUse hook from archmap memory to your project's
.claude/settings.json. It isn't enabled by default because it writes a
file into your project.
Other agents (Cursor, Codex, OpenCode, …)¶
This installs the same skill. Register the MCP server in your agent's MCP
config (see archmap mcp); without it, the skill falls back to
the equivalent CLI commands.
Layout¶
.claude-plugin/marketplace.json # marketplace index (repo root)
plugins/archmap/
├── .claude-plugin/plugin.json # manifest; version tracks the package
├── .mcp.json # archmap MCP server
├── hooks/hooks.json # SessionStart memory injection
└── skills/archmap/SKILL.md # agent workflow
tests/test_agent_plugin.py keeps the plugin version equal to the package
version and fails if the skill names an MCP tool the server doesn't expose.