Skip to content

Repository files navigation

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:

  • Speed and stability correlate — they don't trade off. Forsgren, Humble & Kim's Accelerate (2018) found this, and DORA's 2025 State of DevOps Report replicated it: teams that ship fast are also the teams whose systems stay stable. Under that model, quality gates (tests, linting, type checks, review) are throughput enablers, not a tax on speed — which is why this framework treats them as non-negotiable rather than optional.
  • AI amplifies whatever engineering discipline already exists. DORA's 2025 report found AI adoption alone doesn't predict outcomes; it magnifies the discipline — or its absence — that was already there. This framework's rules, skills, and quality gates exist to give that amplification something good to amplify.
  • The anti-pattern this framework designs against: big-batch, under-reviewed AI changesets. It's the mechanism DORA, GitClear, and the emerging agent-generated-PR research all converge on independently — large, unreviewed diffs are where speed and stability actually start trading off. Swarm orchestration's small-task decomposition and mandatory multi-perspective review exist specifically to keep AI-authored changes out of that failure mode.

What You Get

Commands

Single-agent expert modes, invoked via slash commands, backed by skills in .claude/skills/:

Command Role
/architect System design, ADRs
/builder Implementation, debugging, testing
/qa-engineer Test strategy, E2E, accessibility
/security-auditor Threat modeling, security audits
/ui-ux-designer Interface design, visual assets
/code-check SOLID, DRY, consistency audit

Swarm Orchestrators

Multi-agent commands that fan work out across parallel workers:

Command What It Does
/swarm-plan Launches 3-6 explorer agents to research patterns, dependencies, and constraints — produces a decomposed plan
/swarm-execute Picks up planned work, fans out across builder agents (up to 8 parallel), each running quality gates
/swarm-review Launches 5 parallel reviewers (security, performance, architecture, tests, quality) — run 2-3 times
/swarm-research Deep multi-source investigation with verification tiers

All 12 workflow skills — /architect, /builder, /qa-engineer, /security-auditor, /ui-ux-designer, /code-check, /land-the-plane, /tailor, /swarm-plan, /swarm-execute, /swarm-review, and /swarm-research — carry disable-model-invocation: true: only a user typing the slash name can invoke them. The four role skills (/architect, /qa-engineer, /security-auditor, /ui-ux-designer) are user-invoked entry points: /architect delegates methodology to the always-on designing-systems and writing-adrs library skills, and the others carry their procedures inline or lean on the always-loaded rules (security.md, debugging-protocol.md) — so gating them costs nothing that auto-discovery needed.

The Full Cycle

/architect <feature>  →  /swarm-plan  →  /swarm-execute  →  /swarm-review (2-3x)  →  PR

One agent thinks. Many agents build. Many agents review.

Workers

Five specialized agent types tuned for cost and capability:

Worker Use
worker-explorer Fast codebase search, web research, dependency mapping
worker-builder Implementation, testing, refactoring
worker-reviewer Code review, security analysis
worker-research Deep multi-source investigation
worker-architect Complex design decisions, ADRs

Model tiers are pinned in each agent's frontmatter (.claude/agents/) — that is the single source of truth.

Skills

9 library skills across 5 categories — discovered natively from each skill's description, no hook or registry required:

Architecture · Core Engineering · Operations · Product · Security

A deliberately lean catalog: high-value, single-responsibility skills that don't duplicate what the model already knows, from designing-systems and testing to swarm-coordination and threat-modeling. Generic-knowledge skills (API textbook patterns, OWASP lists, WCAG tables) were retired in the 2026-07 rationalization after base-model evals showed current models produce that content unaided — evidence in artifacts/evals_catalog_rationalization.md. See docs/skills.md for the full list.

Catalog size is a defended design decision, not an oversight. Every skill's name and description loads into every session's context regardless of relevance, and the skill-listing context budget is shared across everything an adopter has installed — this framework's skills plus their own. Growing the catalog without discipline degrades discovery for every skill sharing that budget, including ones this repo didn't add. New skill proposals go through CONTRIBUTING.md's eval-first bar, not "this seems useful."

Measured 2026-07: after gating the 12 workflow skills (disable-model-invocation: true), only the 9 library skills are ungated and auto-discoverable, at roughly ~0.9k tokens of always-loaded listing — against a ~2k-token listing budget on 200K-context sessions (~10k on 1M-context sessions). Separately, CLAUDE.md (104 lines) plus .claude/rules/ (409 lines) run roughly ~5k tokens of always-loaded instructions — a different budget line item from the skill listing. Both numbers move as the platform and catalog change; run /doctor to check for dropped or truncated skill descriptions and /context to see live context-window consumption on your own setup rather than trusting a static figure.

Safety Hooks

Pre-configured hooks that run automatically:

  • Secret detection — blocks commits containing API keys, tokens, private keys
  • Protected files — prevents accidental modification of .env, .mcp.json
  • Push blocking — stops direct pushes to main/master
  • Dangerous command guard — warns on rm -rf, force push, terraform destroy
  • File locking — prevents concurrent edits in multi-agent swarms

What ships enabled: format, warn, and guard hooks, all fail-soft (a missing jq or unparseable input skips the check rather than blocking). Secret-bearing paths and destructive commands are denied at the permission layer (permissions.deny), not just warned about by a hook. See docs/hooks.md#security-model for what hooks can and cannot guarantee.

MCP Servers

Four servers pre-configured in .mcp.json:

Server Purpose
Sequential Thinking Structured multi-step reasoning
Chrome DevTools Browser testing, performance profiling
Context7 Up-to-date library documentation
Filesystem File operations beyond workspace

Customization

Everything is designed to be extended:

  • Add command-style skills → .claude/skills/your-skill/SKILL.md (add disable-model-invocation: true for side-effecting workflows)
  • Add skills → .claude/skills/category/your-skill/SKILL.md
  • Add rules → .claude/rules/your-rule.md
  • Add hooks → .claude/hooks/your-hook.sh
  • Add workers → .claude/agents/worker-yourtype.md

Templates for each are in .claude/templates/.

See docs/customization.md for details.

Docs

About

A More Effective Agent Harness for Claude

Topics

Resources

Contributing

Stars

111 stars

Watchers

2 watching

Forks

Releases

Contributors

Languages