Meta-orchestrator for AI coding agents
One skill to route them all. Works on Claude Code, Codex, Cursor, Windsurf, Antigravity, Gemini CLI, and 5 more platforms.
Quick Start • Platforms • How It Works • Routing Table • Commands • Architecture
You have 40+ Claude Code skills installed. For every task, you manually decide:
- Which skill to invoke?
- Should I combine two skills?
- Does this need the workflow gate (brainstorming) or can I skip it?
- Is caveman mode on? Should I use the caveman variant?
AgentMaster decides for you. One entry point. Zero manual routing.
| Platform | Format | Install Method |
|---|---|---|
| Claude Code | SKILL.md |
bash install.sh or .\install.ps1 |
| OpenAI Codex | SKILL.md |
bash scripts/install-platform.sh --platform codex |
| Cursor | .mdc rules |
bash scripts/install-platform.sh --platform cursor |
| Windsurf | SKILL.md (trigger frontmatter) |
bash scripts/install-platform.sh --platform windsurf |
| Cline | .md rules |
bash scripts/install-platform.sh --platform cline |
| Gemini CLI | SKILL.md |
bash scripts/install-platform.sh --platform gemini |
| Antigravity | SKILL.md (risk/source fields) |
bash scripts/install-platform.sh --platform antigravity |
| OpenCode | SKILL.md (compatibility field) |
bash scripts/install-platform.sh --platform opencode |
| Augment | .md rules |
bash scripts/install-platform.sh --platform augment |
| Aider | CONVENTIONS.md |
bash scripts/convert.sh --tool aider |
| GitHub Copilot | copilot-instructions.md |
bash scripts/convert.sh --tool copilot |
Auto-detect all platforms:
bash scripts/install-platform.sh
# Detects which platforms are installed and copies skills to eachWindows (PowerShell):
git clone https://github.com/Surya8991/AgentMaster.git
cd AgentMaster
.\install.ps1macOS / Linux / WSL:
git clone https://github.com/Surya8991/AgentMaster.git
cd AgentMaster
bash install.shgit clone https://github.com/Surya8991/AgentMaster.git
cd AgentMaster
bash scripts/install-platform.sh --platform codex # or cursor, windsurf, etc.bash scripts/install-platform.sh --platform allThe installer clones 5 dependency repos, converts skills to platform-specific formats, and skips anything already installed.
Restart your IDE/agent after installation.
You say something. AgentMaster classifies it and routes to the right skill(s):
You: "Build an auth system with JWT"
AgentMaster: Routing → brainstorming (workflow) + wshobson backend-development/security-scanning (auto-surface)
You: "Fix this crash in the payment module"
AgentMaster: Routing → systematic-debugging
You: "Write SEO copy for the landing page"
AgentMaster: Routing → anthropic-skills:running-marketing (+ copywriting, ai-seo)
You: "Deploy to AWS with Terraform"
AgentMaster: Routing → writing-plans (complex infra) + devops
You: "How did we solve that auth bug last week?"
AgentMaster: Routing → mem-search (session memory)
No manual skill selection. No guessing. No loading 5 skills when you need 1.
| Command | What It Does |
|---|---|
/agent-master |
Invoke on current task — auto-classifies and routes |
/agent-master route <task> |
Dry-run — shows routing plan without executing |
/agent-master status |
Shows current state (caveman on/off, last skill used, active workflow) |
/agent-master update |
Force-update all dependency repos and sync skills |
/agent-master doctor |
Health-check installed skills, caches, and pins (PASS/WARN/FAIL) |
/agent-master list |
List installed skills grouped by source repo |
/agent-master routes |
Routing log, learned overrides, and unrouted skills |
/agent-master routes analyze |
Rank repeated misroutes and busiest categories |
/agent-master scaffold [apply] |
Report (or create) the canonical project-doc spine — README/AGENTS.md/CLAUDE.md/PROJECTLOG.md |
/agent-master profile [name] |
Show or switch the install profile (full/dev/business/minimal) |
/agent-master rollback [repo] |
Restore a repo's skills to the pre-sync version and pin it |
/caveman |
Enable token compression (~75% savings on all output) |
/codereview |
Blunt, factual code review — finds bugs, security issues, doc mismatches |
AgentMaster automatically pulls latest versions from all dependency repos on first invoke each session (background, non-blocking). Has a 6-hour cooldown to avoid spamming GitHub. Use /agent-master update to force an immediate sync.
Dependency repos are declared in repos.manifest (one line per repo: name|url|skill_source). Add personal repos in ~/.claude/.agentmaster-cache/repos.local (same format) — they survive self-updates. Pin any repo to an exact commit via repos.pins (name=sha). Every sync writes a report to ~/.claude/.agentmaster-cache/last-sync-report.txt including before→after commits and skill-name collisions between repos.
Don't want all ~195 skills? Install a subset: bash install.sh --profile dev (or .\install.ps1 -Profile dev). Shipped profiles in profiles.manifest: dev (compression + superpowers workflow + session memory + frontend design + wshobson/davila7 tactical + all custom skills), business (compression + orchestrator + repo packing), minimal (orchestrator + compression), and the default full (everything). Switch anytime with /agent-master profile <name> — excluded skills are pruned (only ones AgentMaster installed; your own skills are never touched) and the subset resyncs. Profiles persist across auto-updates; define your own in ~/.claude/.agentmaster-cache/profiles.local. The orchestrator skill itself is always installed, and an unknown profile name falls back to full rather than uninstalling anything.
Every sync that changes a repo automatically backs up the outgoing skill versions first (~/.claude/.agentmaster-cache/backups/<repo>/, one generation each). If an upstream update breaks something, /agent-master rollback <repo> restores the previous versions and pins the repo in pins.local so auto-updates hold the rollback instead of re-clobbering it 6 hours later. Works for the orchestrator itself too (rollback agent-master). Undo by deleting the repo's line from pins.local — the next sync resumes tracking upstream. pins.local also serves as a general machine-local pin file: it wins over the repo-committed repos.pins.
Routing learns from use. Every routing decision is logged to ~/.claude/.agentmaster-cache/routing-log.txt. When you correct a misroute in-session ("no, use X"), the correction is saved as a rule in routing-overrides.md, which loads every session and takes precedence over the static routing table — so the same mistake isn't made twice. Skills that sync in but aren't in the routing table are detected on every update (unrouted-skills.txt) and consulted as a fallback before AgentMaster asks you what a task is. Inspect all three with /agent-master routes.
/agent-master route deploy HIPAA-compliant API to AWS
Output:
AgentMaster Route Plan
━━━━━━━━━━━━━━━━━━━━━
Task: deploy HIPAA-compliant API to AWS
Category: DevOps/Deploy + Compliance
Workflow: writing-plans (complex infra)
Domain: devops + anthropic-skills:compliance-handling
Entry point: writing-plans
Combination: devops + compliance-handling (HIPAA — flag regulated depth is out of scope)
Conflicts: none
Three layers that stack, never compete:
┌─────────────────────────────────────────────┐
│ OUTPUT LAYER: caveman │
│ Token compression. Active when user │
│ enables /caveman. Layers on top of ANY │
│ other skill output. │
├─────────────────────────────────────────────┤
│ WORKFLOW LAYER: superpowers │
│ brainstorm → plan → TDD → review → finish │
│ Active for code/engineering tasks ONLY. │
│ Non-negotiable: can't skip brainstorming. │
├─────────────────────────────────────────────┤
│ DOMAIN LAYER: anthropic-skills (business) │
│ + wshobson/davila7/impeccable (technical) │
│ + custom (devops, security-audit). │
│ ~195 installed skills provide expertise. │
└─────────────────────────────────────────────┘
Key principle: AgentMaster names the most-specific entry skill; related tactical/domain skills surface on their own by description (they are individual skills, not bundles that fan out).
| Category | Signal Words | Routes To |
|---|---|---|
| Build/Create | build, create, implement, scaffold | brainstorming → superpowers + wshobson backend-development/python-development, davila7 database (auto-surface) |
| Refactor | refactor, restructure, clean up | brainstorming → wshobson python-anti-patterns/python-design-patterns/architecture-patterns |
| Debug/Fix | bug, crash, error, fix, broken | systematic-debugging + wshobson python-error-handling/python-resilience |
| Code Review | review code, /codereview, check diff | codereview (blunt) or requesting-code-review (PR workflow) |
| Commit/Ship | commit, merge, ship, push | finishing-a-development-branch |
| Test | write tests, TDD, coverage | test-driven-development + wshobson python-testing-patterns, anthropic-skills webapp-testing |
| Marketing | blog, SEO, campaign, ads, copy | anthropic-skills:running-marketing (+ cold-email, copywriting, ai-seo) |
| Strategy | fundraise, roadmap, pivot, board | anthropic-skills:gtm-foundations (+ crafting-positioning) |
| Product | PRD, user stories, personas | anthropic-skills:building-product (+ validating-customers) |
| Finance | DCF, budget, runway, ARR | anthropic-skills:setting-pricing + xlsx (no dedicated valuation skill — flagged) |
| Business Growth | churn, pipeline, RFP, proposal | anthropic-skills:executing-sales (+ pipeline-management) |
| Project Mgmt | Jira, sprint, scrum, retro | anthropic-skills:schedule + writing-plans (thin coverage — flagged) |
| Compliance | ISO, FDA, GDPR, SOC2, audit | anthropic-skills:compliance-handling + security-audit (regulated depth flagged) |
| Database Design | schema, migration, query optimization, SQL, Postgres | davila7 database (sql-pro, database-optimizer, postgresql-optimization) |
| DevOps/Deploy | Docker, CI/CD, Terraform, AWS, Kubernetes | devops + wshobson cicd-automation/kubernetes-operations |
| Security | OWASP, XSS, vulnerability, pen test | security-audit + wshobson security-scanning (STRIDE, SAST) |
| UI/UX Design | design/polish/audit UI, wireframe, color, layout, animation | impeccable (frontend craft) or anthropic-skills:ui-ux-pro-max (reference) |
| Documentation | write docs, generate PDF/DOCX | anthropic-skills:docx/pdf/pptx/xlsx |
| Research | analyze market, deep dive, investigate | anthropic-skills:deep-research |
| LLM/AI App Dev | RAG, embeddings, LangChain, prompt engineering, MCP | brainstorming + wshobson llm-application-dev, anthropic-skills:mcp-builder |
| Memory/History | last time, previous session, how did we | mem-search / timeline-report |
| Explore Codebase | code structure, find functions | smart-explore |
| Whole-Codebase Analysis | entire repo, full audit, architecture review | repomix-pack → analysis skill |
| Task Dashboard | task board, kanban, visualize tasks | task-viewer (launches claude-task-viewer) |
| Simple Question | factual question, no action | direct answer (no routing) |
When words match multiple categories, AgentMaster resolves automatically:
| Ambiguous Word | Default | Override When... |
|---|---|---|
| "pipeline" | DevOps | "sales pipeline" → Business Growth |
| "design" | UI/UX | "system design" → Build/Create |
| "test" | Code Test | "A/B test" → Marketing |
| "audit" | Compliance | "security audit" → Security |
| "container" | DevOps | "container component" → Build/Create |
| "review" | Code Review (codereview) |
"review PR" → requesting-code-review. "content review" → Marketing |
| "deploy" | DevOps | "deploy feature" + code → Build + DevOps |
| "report" | Documentation | "bug report" → Debug/Fix |
| "plan" | Build | "make a plan" → make-plan |
| Combo | Routing Strategy |
|---|---|
| Code + DevOps | Superpowers workflow + devops for deployment |
| Security + Code | security-audit for findings + superpowers for fixes |
| Security + Compliance | security-audit + anthropic-skills:compliance-handling |
| Security + DevOps | security-audit (app) + devops (infra) |
| Product + Code | anthropic-skills:building-product for requirements, then superpowers for code |
| UI/UX + Code | impeccable for design + implementation (writes working code directly); ui-ux-pro-max for reference-only |
| Research + Any | deep-research first, then domain skill |
| Code + Docs | Build first, then anthropic-skills:docx/pdf |
Hard limit: Maximum 2 domain skills per request. If 3+ detected, AgentMaster asks you to narrow scope.
| Source | What | Skills |
|---|---|---|
| This repo | Orchestrator + custom skills | agent-master, devops, security-audit, codereview, repomix-pack, task-viewer |
| caveman | Token compression | caveman, caveman-commit, caveman-review, caveman-help, compress |
| superpowers | Dev workflow | brainstorming, writing-plans, test-driven-development, systematic-debugging, +10 more |
| claude-skills | Repeatable agent loops | loop-library (repo is a plugin-of-plugins; its business bundles don't install flat — business domains are served by the anthropic-skills plugin instead) |
| claude-mem | Session memory | mem-search, smart-explore, knowledge-agent, make-plan, do, timeline-report, version-bump |
| impeccable | Frontend design | impeccable (design/polish/audit + 23 commands) |
| anthropics/skills | Official Anthropic skills | docx, pdf, pptx, xlsx, frontend-design, mcp-builder, skill-creator, webapp-testing, +9 more |
| wshobson/agents | Technical tactical (15 curated subsets) | backend-development, python-development, kubernetes-operations, cicd-automation, security-scanning, llm-application-dev, javascript-typescript, frontend-mobile-development, ui-design, cloud-infrastructure, observability-monitoring, incident-response, developer-essentials, data-engineering, documentation-generation — 96 skills |
| davila7/claude-code-templates | DB + git tactical (2 curated subsets) | database, git — 16 skills |
| AgentMaster (self-update) | Keeps orchestrator current | auto-synced via ~/.claude/.agentmaster-cache/agent-master |
| Total | ~195 skills (full profile) |
repomix-pack wraps repomix so any skill can analyze the entire repo cheaply.
Auto-runs at the start of every new session when AgentMaster detects a code repo (.git/, package.json, pyproject.toml, Cargo.toml, or go.mod present). The snapshot lives at .agentmaster/codebase.xml and is reused by security-audit, codereview, and refactor passes for whole-repo analysis. Has a built-in staleness check so it only re-packs when source files change.
# Manual control
/agent-master repomix # pack now
/agent-master repomix refresh # force re-pack
/agent-master repomix include src/**
"skip repomix" # disable for this sessionRequires npm install -g repomix (the installer does this automatically).
Note on Mem0: Mem0 was evaluated and intentionally not added —
claude-mem(already bundled viamem-search,knowledge-agent,timeline-report) covers the same persistent-memory use case locally without an extra API key or cloud dependency.
| Rule | What Happens |
|---|---|
| No self-invocation | AgentMaster cannot call itself |
| Max depth = 2 | AgentMaster → Skill → Sub-skill. Third hop blocked. |
| No circular calls | If Skill A triggered AgentMaster, it won't call Skill A back |
| When blocked | States assumption clearly, returns to user |
When /caveman is active, AgentMaster automatically:
- Compresses all its own output (drop articles, fragments OK)
- Uses caveman variants when available (
caveman-reviewinstead ofrequesting-code-review) - Never lets caveman mode block or modify routing decisions
| Metric | Without AgentMaster | With AgentMaster |
|---|---|---|
| Token usage | Baseline (100%) | ~30% (70% savings) |
| Task accuracy | 85% | 90% |
| Speed | 1x | 0.9x (10% routing overhead) |
| Manual effort | High (pick skills yourself) | None (auto-routes) |
AgentMaster/
├── skills/ # Source of truth (Claude Code format)
│ ├── agent-master/SKILL.md # Meta-orchestrator (26 categories)
│ ├── codereview/SKILL.md # Blunt code review (/codereview)
│ ├── devops/SKILL.md # CI/CD, Docker, Terraform, deployment
│ ├── repomix-pack/SKILL.md # Whole-codebase snapshot via repomix
│ ├── security-audit/SKILL.md # OWASP Top 10, vuln scanning
│ └── task-viewer/SKILL.md # Launches claude-task-viewer dashboard
│
├── .cursor/rules/ # Auto-generated Cursor format
├── .windsurf/skills/ # Auto-generated Windsurf format
├── .clinerules/ # Auto-generated Cline format
├── .codex/skills/ # Auto-generated Codex format
├── .gemini/skills/ # Auto-generated Gemini CLI format
├── .antigravity/skills/ # Auto-generated Antigravity format
├── .opencode/skills/ # Auto-generated OpenCode format
├── .augment/rules/ # Auto-generated Augment format
├── .github/copilot-instructions.md # Auto-generated Copilot format
├── CONVENTIONS.md # Auto-generated Aider format
│
├── scripts/
│ ├── convert.sh # Generate platform files from skills/
│ └── install-platform.sh # Install to any platform's skill directory
│
├── Tool-Stack-Reference/ # 1524 tools + 1767 tech items (free-first, no browser ext.)
│ ├── hub/ # Auto-generated from Master Dev Hub (committed)
│ │ ├── README.md # Index of all 62 hub reference files
│ │ ├── tools-ai-coding.md # 143 AI coding tools (free-first)
│ │ ├── tools-ai-agents.md # 53 AI agents & automation tools
│ │ ├── tools-ai-infra.md # 42 AI infrastructure tools
│ │ ├── tools-api-dev-tools.md # 49 API & dev tools
│ │ ├── ... # 28 more tool category files
│ │ ├── tech-js-ts-packages.md # 210 JS/TS packages
│ │ ├── tech-python-packages.md # 200 Python packages
│ │ └── ... # 27 more tech category files
│ ├── scripts/ # Regeneration scripts (run when hub HTML updates)
│ │ ├── extract.js # Parse master-hub.html → JSON intermediates
│ │ └── generate-md.js # JSON intermediates → hub/*.md files
│ ├── README.md # Workflow: how to use + how to regenerate
│ └── index.html # Searchable HTML dashboard
│
├── agents.md # Agent context file (read first every session)
├── RULES.md # Active project rules — loaded by AgentMaster on session start
├── RULES.example.md # Instructional template with placeholders (copy + customize)
├── install.sh # Claude Code full installer (+ dependencies)
├── install.ps1 # Claude Code full installer (Windows)
└── uninstall.sh # Remove custom skills only
Edit skills in skills/ → run bash scripts/convert.sh --tool all → all platform files regenerate.
# Remove only AgentMaster custom skills (agent-master, codereview, devops, security-audit, repomix-pack)
# Keeps third-party skills (caveman, superpowers, claude-skills, claude-mem, impeccable, anthropic-official, wshobson, davila7) intact
bash uninstall.sh
# Nuclear option: remove everything
rm -rf ~/.claude/skills/*Two rules files — one active, one template:
| File | Purpose |
|---|---|
RULES.md |
Active rules — loaded by AgentMaster at session start when present in repo root |
RULES.example.md |
Template — instructional copy with placeholders, copy and customize |
AgentMaster reads RULES.md from the repo root on first invocation each session. Rules defined there govern commit style, output style, paths, workflow, and agent behavior for that session.
# Set up global rules (applies to all sessions)
cp RULES.example.md ~/.claude/RULES.md
# Edit: replace [YOUR_GITHUB_USERNAME], [YOUR_PROJECTS_DIR], etc.
# Project-level rules (applies only in this repo)
# Already present as RULES.md — edit directlySee RULES.example.md for the full annotated template.
- Edit SKILL.md files in
skills/(source of truth) - Run
bash scripts/convert.sh --tool all(orpowershell -File scripts\convert.ps1 -All) to regenerate platform files - Test with
/agent-master route <your task>(dry-run mode) - Run
bash tests/smoke.sh(CI runs this +shellcheck --severity=error) - Commit and PR
The runtime loads skills from ~/.claude/skills/, and the /agent-master subcommands run
from the self-update cache — both are normally populated by git pull from the remote, so a
local edit has no effect until you commit, push, and let the auto-updater sync. To iterate
without that round-trip:
bash scripts/dev-sync.sh # push your working tree straight to the live install
# ...edit skills/scripts, re-run dev-sync.sh, test in a new session...
bash scripts/dev-sync.sh --off # resume normal auto-update from remotedev-sync writes a .dev marker that stops the auto-updater from overwriting your local
edits. /agent-master doctor reports when dev mode is on. (Windows: scripts\dev-sync.ps1.)
AgentMaster is built on top of these excellent open-source projects:
| Project | Author | What It Provides |
|---|---|---|
| caveman | Julius Brussee | Token compression (~75% savings), commit/review variants |
| superpowers | Jesse Vincent | Dev workflow (brainstorming, TDD, debugging, code review) |
| claude-skills | Alireza Rezvani | 235+ domain skills (engineering, marketing, product, C-level, finance) |
| claude-mem | thedotmack | Persistent session memory, smart-explore, knowledge agents |
AgentMaster does not bundle or redistribute their code — the installer clones them at install time. All credit for those skills belongs to their respective authors.
MIT
Built by Surya L with Claude Code