A Workflow OS for AI Code Editors
Persistent memory · 4-layer skill mesh · Privacy hooks · Observability dashboard
Install • Features • Architecture • Docs • Contributing
UltraThink turns AI code editors from stateless assistants into persistent, skill-aware agents that remember your preferences, enforce your standards, and adapt to your workflow.
You ──► AI Editor ──► UltraThink ──► Skills matched · Memories recalled
Context injected · Better responses
Copy this prompt into Claude Code, Cursor, Windsurf, or any AI editor and it will handle everything:
Copy this prompt
I want to install UltraThink — a Workflow OS that gives you persistent memory,
auto-triggered skills, privacy hooks, and an observability dashboard.
Steps:
1. Clone the repo: git clone https://github.com/InugamiDev/ultrathink-oss.git ~/ultrathink
2. cd ~/ultrathink
3. Run ./scripts/setup.sh — this installs all dependencies (root, dashboard, memory),
creates .env from the template, makes hooks executable, and verifies skill files.
4. Run ./scripts/init-global.sh — this symlinks skills, agents, references, and hooks
into ~/.claude/ so every Claude Code session has UltraThink capabilities.
5. If we're in Cursor, Windsurf, Antigravity, or Copilot, also run
./scripts/sync-editors.sh --all to generate editor-specific config files.
6. After setup, tell me:
- How many skills were linked
- How many hooks were installed
- Whether .env needs a DATABASE_URL (for memory persistence)
- The URL for the dashboard (should be http://localhost:3333)
For the database: I need a free Neon Postgres account from https://neon.tech.
Once I have a DATABASE_URL, put it in ~/ultrathink/.env and run:
cd ~/ultrathink/memory && npx tsx scripts/migrate.ts
To start the dashboard: npm run dashboard:dev from ~/ultrathink/
Show me the final status when done.
git clone https://github.com/InugamiDev/ultrathink-oss.git ~/ultrathink
cd ~/ultrathink && ./scripts/setup.sh && ./scripts/init-global.shThen run claude in any project. Full guide →
| Editor | Integration | Guide |
|---|---|---|
| Claude Code | Full (hooks, skills, memory, auto-trigger) | install-claude-code.md |
| Cursor | Rules (.cursor/rules/) |
install-cursor.md |
| Windsurf | Rules (.windsurf/rules/) |
install-windsurf.md |
| Antigravity | Rules (GEMINI.md) |
install-antigravity.md |
| GitHub Copilot | Rules (.github/copilot-instructions.md) |
install-copilot.md |
| OpenClaw | Skills + MCP servers | install-openclaw.md |
# Generate configs for all editors at once
./scripts/sync-editors.sh --all45 active skills across 4 layers, auto-triggered by intent detection on every prompt (<30ms).
| Layer | Examples | Purpose |
|---|---|---|
| Orchestrator | cook, ship, gsd-verify |
End-to-end workflows |
| Hub | debug, test, react |
Multi-step coordinators |
| Utility | fix, refactor, verify |
Focused tools |
| Domain | nextjs, tailwind, react |
Tech specialists |
340+ more in _archive/ — restore any with mv. Skills catalog →
Postgres-backed persistent memory with 3-tier search (tsvector + trigram + ILIKE). Memories are scoped by project, ranked by importance, and recalled automatically at session start.
Block .env, .pem, credentials, and secrets before the model sees them. All decisions logged.
Next.js 15 on port 3333 — memory browser, skill graph, hook events, usage tracking, kanban, plans.
3-line CLI status bar: model, context %, quotas, active skills, memory count, Tekiō spins.
┌─────────────────────────────────────────────────┐
│ Claude Code CLI │
├─────────────────────────────────────────────────┤
│ │
│ SessionStart → memory recall + statusline │
│ PromptSubmit → skill scoring → top 5 injected │
│ PreToolUse → privacy check (block creds) │
│ PostToolUse → quality gate + observe + save │
│ Stop → memory flush + session close │
│ │
│ ┌────────────────────────────────────────────┐ │
│ │ Neon Postgres (pgvector + pg_trgm) │ │
│ │ memories · sessions · hooks · skills │ │
│ │ plans · tasks · decisions · adaptations │ │
│ └────────────────────────────────────────────┘ │
│ │
│ ┌────────────────────────────────────────────┐ │
│ │ Skill Mesh (4 layers, 45 active) │ │
│ │ Orchestrators → Hubs → Utils → Domain │ │
│ │ Auto-trigger + graph traversal (linksTo) │ │
│ └────────────────────────────────────────────┘ │
│ │
│ ┌────────────────────────────────────────────┐ │
│ │ Dashboard (Next.js 15, :3333) │ │
│ │ /memory /skills /hooks /usage /kanban │ │
│ └────────────────────────────────────────────┘ │
└─────────────────────────────────────────────────┘
ultrathink/
├── .claude/
│ ├── hooks/ 20 lifecycle hooks
│ ├── skills/ 45 active + _archive/ (340+)
│ ├── agents/ 10 agent definitions
│ ├── references/ Quality, privacy, teaching rules
│ └── commands/ Slash commands
├── memory/ Postgres-backed memory system
├── dashboard/ Next.js 15 observability UI
├── openclaw/ OpenClaw skill package + MCP config
├── tests/ Vitest test suite
├── scripts/ Setup + utilities
└── docs/ Installation + reference docs
| Topic | File |
|---|---|
| Install: Claude Code | install-claude-code.md |
| Install: Cursor | install-cursor.md |
| Install: Windsurf | install-windsurf.md |
| Install: Antigravity | install-antigravity.md |
| Install: GitHub Copilot | install-copilot.md |
| Install: OpenClaw | install-openclaw.md |
| Editor setup details | editor-setup.md |
| Skills catalog | skills-catalog.md |
| Memory system | memory-system.md |
| Hooks & privacy | hooks-and-privacy.md |
| Dashboard overview | dashboard-overview.md |
| Creating skills | how-to-create-a-new-skill.md |
| Linking skills | how-to-link-skills.md |
| Database schema | memory-schema.md |
| Troubleshooting | troubleshooting.md |
| Font pairings (230+) | font-pairings.md |
| Design languages & UX laws | design-languages-ux-laws.md |
| Themes & design systems | design-themes-and-systems.md |
./scripts/setup.sh # Full setup
./scripts/init-global.sh # Install into ~/.claude/
./scripts/sync-editors.sh --all # Generate all editor configs
npm run dashboard:dev # Dashboard on :3333
npm run test # Run testsSee CONTRIBUTING.md. Quick start:
git clone https://github.com/InugamiDev/ultrathink-oss.git
cd ultrathink-oss && ./scripts/setup.sh && npm run testUltraThink is an integration project, not an innovation project. Most concepts come from the community:
| Concept | Origin |
|---|---|
| TDD, brainstorming, subagent workflow | Superpowers (121K+ stars) |
| Multi-agent + hook architecture | Everything Claude Code (115K+ stars) |
| SKILL.md format | Anthropic Skills (official) |
| Multi-layer security hooks | Claude Forge |
| "Second brain" memory pattern | coleam00/second-brain-skills |
| Hook lifecycle patterns | claude-code-hooks-mastery |
What UltraThink adds on top:
- Database-backed memory — Real Postgres (pgvector + pg_trgm) instead of flat MEMORY.md files. Memories persist across sessions with importance ranking, fuzzy search, and synonym expansion.
- Observability dashboard — A full Next.js 15 UI (skill graph, memory browser, hook events, usage tracking) that no other repo provides at this level.
- Auto-trigger engine — Every prompt is scored against the skill registry in <30ms. Top 5 skills are injected automatically via intent detection + graph traversal. Neither Superpowers nor ECC do this.
- Cross-editor configs — One command generates rules for Cursor, Windsurf, Antigravity, Copilot, and OpenClaw.
- Adaptive learning (Tekiō) — Failures become immunity rules, successes get reinforced. Infinite wheel spins.
Where others are stronger:
- Superpowers has a tighter methodology (spec → plan → TDD → review) and a massive community.
- Everything Claude Code has more agents (28), more skills (126), 98% test coverage, and 115K+ stars.
- Both are battle-tested by thousands of developers. UltraThink is a personal workflow open-sourced.
Bottom line: If you want a proven, community-backed framework, use Superpowers or ECC. If you want database-backed memory, a visual dashboard, and auto-triggered skills wired together, UltraThink fills that gap.
| Project | Author | Integration |
|---|---|---|
| Superpowers | obra | TDD, debugging, brainstorming, plan execution |
| Everything Claude Code | affaan-m | Multi-agent architecture, hook patterns |
| Claude Forge | sangrokjung | Multi-layer security hooks |
| Second Brain Skills | coleam00 | "Second brain" framing for Claude Code |
| claude-code-hooks-mastery | disler | Hook lifecycle patterns |
| VFS | TrNgTien | AST-based token compression (60-98% savings) |
| Get Shit Done | gsd-build | Spec-driven planning, wave execution |
| Impeccable | pbakaus | Frontend design skill suite |
| Anthropic Skills | Anthropic | Skill format conventions |
MIT — see LICENSE.
Built by InuVerse