Claude Agentic Framework

Claude Agentic Framework

A drop-in template for Claude Code projects. Adds coordinated multi-agent swarms, specialized workflow skills, a lean core of reusable library skills, and safety hooks — all configured through a single install command.

Two ways to adopt

(a) Raw drop-in — recommended, full-featured

Clone or copy the repo's files directly into your project. This is the only path that gives you everything: skills, agents, hooks, .claude/settings.json permission rules (including the permissions.deny secret-file guards), and .claude/rules/ (tech strategy, security standards, core directives).

git clone https://github.com/dralgorhythm/claude-agentic-framework.git
cd your-project
../claude-agentic-framework/scripts/init-framework.sh .

To pin to a known-good release: git checkout vX.Y.Z before running the script.

(Clone-then-run, deliberately — this framework's own permission rules deny pipe-to-shell installs, and its docs tell you to review scripts before executing them. Practice what we ship.)

The script will:

• Copy .claude/ (skills, rules, hooks, agents, templates)
• Copy .mcp.json (MCP server configuration)
• Copy CLAUDE.md and AGENTS.md (project instructions)
• Create an artifacts/ directory for planning documents
• Set up .gitignore entries

The script prompts before overwriting any existing files. To pull in framework updates, cd into your clone, git pull, then re-run the script. The framework is zero-install — no dependencies to fetch. New versions are announced on the GitHub Releases page.

After install:

1. Edit CLAUDE.md — Add your build/test commands and project context
2. Edit .claude/rules/tech-strategy.md — Configure your tech stack (this is required — the framework enforces whatever you put here)
3. Start Claude Code and try: /architect hello

(b) Plugin install — skills, agents, and hooks only

Install directly inside a Claude Code session, no cloning required:

/plugin marketplace add dralgorhythm/claude-agentic-framework
/plugin install agentic-framework@agentic-framework

This gets you the skills, the five worker agents, and the guardrail hooks, wired up automatically. It does not replace cloning the repo — it's a lighter-weight path with real gaps:

• Skills are namespaced. Invoke them as /agentic-framework:architect, /agentic-framework:builder, etc., not the bare /architect names used in the raw drop-in.
• No .claude/settings.json permission rules ship with the plugin. The permissions.deny guards for secrets, .env*, and other sensitive paths are repo-level configuration — a plugin install will not add them to your project. You get the hooks that warn/guard, but not the deny-layer enforcement described above.
• No .claude/rules/ ship with the plugin. tech-strategy.md, security.md, core-directives.md, agent-constraints.md, and code-quality.md are repo-level files, not plugin content — they won't appear in your project via plugin install.
• Plugin agents ignore permissionMode frontmatter. Any per-agent permission mode set in an agent's frontmatter is not honored when the agent is loaded as a plugin agent.

If you need the full guardrail set (deny rules, rules directory, settings.json), use the raw drop-in instead — the plugin path trades completeness for a faster, in-session install.

Plugin content updates at each tagged release — run /plugin update to pull the latest; after updating, check MIGRATION.md for breaking changes.

For maintainers: validate packaging with claude plugin validate --strict . (marketplace) and claude plugin validate .claude-plugin/plugin.json (plugin — reports one expected warning that the repo-level CLAUDE.md is not loaded as plugin context; that is by design, per the caveats above).

This repo currently ships with no LICENSE file; nothing here should be read as a license grant.

Works beyond Claude Code

.claude/skills/ isn't a Claude-only format anymore. Agent Skills is an open standard now supported by 40+ Agent Skills clients (see agentskills.io) — the GitHub Copilot agent family (cloud agent, code review, CLI, VS Code/JetBrains agent mode) and standalone VS Code all discover and read skills from .claude/skills/ in place, no conversion step required. That means the catalog travels with the repo regardless of which client opens it. Claude Code-specific frontmatter (model, argument-hint, disable-model-invocation, and similar) sits outside the shared spec; non-Claude readers ignore fields they don't recognize rather than failing on them, so the same files degrade gracefully across tools.

Why This Shape

Three findings anchor the design, dated because the evidence base moves: