claude-fulcrum
Advanced tools
The operating system for AI-powered development — 25 agents, 112 skills, 62 commands, 4-signal hybrid memory, knowledge graphs, and cross-platform support for Claude Code, Codex, Cursor, Copilot & OpenCode
26 Agents · 122 Skills · 66 Commands · 9 Language Rulesets · 1,536 Tests · 6 Platforms
One harness. Every AI coding tool. Unified memory. Swarm orchestration. Neural search. Design intelligence.
Quick Start · A Day with Fulcrum · Architecture · Core Systems · Design Intelligence · Agents · Skills · Performance · Roadmap
What is this? Claude Fulcrum is a production-grade agent harness that unifies Claude Code, Codex CLI, Cursor, GitHub Copilot, and OpenCode into a single development environment — with shared vector memory, 4-signal neural search, knowledge graphs, swarm orchestration, 33 lifecycle hooks, 26 specialized AI agents, and a full UI/UX design intelligence suite that activate automatically based on what you're doing.
Claude Code knows what you did in Claude Code. Cursor knows what you did in Cursor. Copilot knows nothing.
None of them share memory. None of them learn from each other. Every session, every tool, starts from zero.
Fulcrum fixes this with one shared memory layer — FTS5 + vector embeddings + knowledge graph — that all 6 platforms read from and write to simultaneously. Patterns learned yesterday are available today. Plans made in one tool execute in another.
You write code in Cursor → Fulcrum's agents review it automatically
You plan a feature in Claude Code → Codex workers execute the plan in parallel
Copilot suggests a completion → It already knows your team's patterns from memory
A build fails in any tool → The build-error-resolver agent fixes it
You fix a tricky bug → The knowledge graph remembers the pattern forever
You need a design system → 7 design skills generate tokens, palettes, and typography instantly
The codebase remembers. So you don't have to.
9:00am — You open Cursor. Before you type a word, Fulcrum has already loaded context from yesterday's Claude Code session. The auth bug you were debugging? The knowledge graph remembers the pattern. Copilot's first suggestion already reflects it.
10:30am — You start a feature. /plan generates a phased implementation with risk analysis. Codex workers execute phases 2 and 3 in parallel worktrees while you handle phase 1 in Claude Code. Three agents working simultaneously on the same feature.
2:00pm — A build breaks. Before you notice, the build-error-resolver agent has already diagnosed it. The quality loop tries three approaches. The third passes all 1,536 checks. It opens the fix for your review.
5:00pm — You need a UI for the feature. Seven design skills activate automatically. BM25 search finds the right palette for your stack. Token architecture generated. shadcn/ui components scaffolded. The design system matches your brand.
5:30pm — Session ends. Everything is stored. Tomorrow's tools already know what today accomplished.
This is not a future workflow. This is Fulcrum today.
git clone https://github.com/ORION2809/claude-Fulcrum.git
cd claude-Fulcrum
npm install
Install for your stack:
# Unix / macOS
./install.sh typescript # or: python, golang, kotlin, cpp, perl, php, swift
# Windows
.\install.ps1 typescript
Verify everything works:
npm test # 1,536 tests — should all pass
/plugin install claude-fulcrum
npx claude-fulcrum typescript
Install any skill (or all of them) to any platform — sequentially or in parallel via swarm:
# Install all skills to all 6 platforms (sequential)
npx claude-fulcrum skill-install --all --platform all
# Install specific skills to specific platforms
npx claude-fulcrum skill-install --skills tdd-workflow,api-design --platform claude,copilot
# Parallel swarm mode — one worker per platform
npx claude-fulcrum skill-swarm --all --platform all
# Preview what would be installed (dry run)
npx claude-fulcrum skill-install --dry-run --all --platform all
# List all available skills
npx claude-fulcrum skill-install --list
GitHub Copilot can consume Claude-format skills directly from .claude/skills/, so Fulcrum skills written for Claude Code can be reused in Copilot without a separate skill format conversion.
| Platform | Scope | Destination |
|---|---|---|
| Claude | Home | ~/.claude/skills/ |
| Cursor | Project | .cursor/skills/ |
| Codex | Home | ~/.codex/skills/ |
| OpenCode | Home | ~/.opencode/skills/ |
| Antigravity | Project | .agent/skills/ |
| Copilot | Project | .claude/skills/ + optional .github/ scaffold |
graph TB
subgraph PLATFORMS["🖥️ Platform Layer"]
CC[Claude Code]
CX[Codex CLI]
CU[Cursor]
CP[Copilot]
OC[OpenCode]
end
subgraph ENGINE["⚙️ Orchestration Engine"]
SW["Swarm Coordinator<br/>claude-flow · 4 topologies"]
TR["Task Router<br/>3-tier model routing"]
QE["Quality Enforcer<br/>lint → test → review → fix"]
end
subgraph MEMORY["🧠 Memory Layer"]
SQ["SQLite + FTS5<br/>Canonical Store"]
VS["Vector Search<br/>384-dim ONNX"]
KG["Knowledge Graph<br/>12 entity types"]
RRF["RRF Fusion<br/>4-signal ranking"]
end
subgraph AGENTS["🤖 26 Agents"]
DEV["Development<br/>planner · architect · tdd"]
REV["Reviewers<br/>code · security · 6 languages"]
OPS["Operations<br/>chief-of-staff · loop-operator"]
end
subgraph DESIGN["🎨 Design Intelligence"]
UX["ui-ux-pro-max<br/>67 styles · BM25 search"]
DS["design-system<br/>3-layer tokens · CSS"]
UI["ui-styling · brand<br/>shadcn · Tailwind · 65+ fonts"]
end
CC & CX & CU & CP & OC --> ENGINE
ENGINE --> MEMORY
ENGINE --> AGENTS
ENGINE --> DESIGN
MEMORY --> CC & CX & CU & CP & OC
Every platform reads from and writes to the same shared context:
| Platform | Role | Reads | Writes |
|---|---|---|---|
| Claude Code | Deep workflows — TDD, planning, security | Everything | Patterns → memory |
| Codex CLI | Fast parallel execution | Plans, tasks | Results, code |
| Cursor | Visual IDE integration | Agents, skills | Code changes |
| GitHub Copilot | Inline completions + chat | Standards, patterns | — |
| OpenCode | Open-source alternative | Full agent/skill library | Code, docs |
Full guide: docs/CROSS_PLATFORM_INTEGRATION.md
Not just a database — a 4-signal neural search system backed by sql.js (WASM SQLite) with 4 schema migrations, 50+ API methods, and 12 validated entity types.
| Signal | How It Works | Role |
|---|---|---|
| Lexical | FTS5 full-text search with porter stemming + unicode tokenization (LIKE fallback) | Keyword precision |
| Recency | Exponential time-decay scoring — recent observations rank higher | Temporal relevance |
| Structure | Category and entity-type boosting from the knowledge graph | Contextual depth |
| Vector | Cosine similarity on 384-dimensional embeddings via Reciprocal Rank Fusion | Semantic matching |
Embedding pipeline: ONNX Runtime primary (Xenova/all-MiniLM-L6-v2, 384-dim) with deterministic hash fallback. Vectors computed automatically on every observation and memory note write.
Search pipeline:
Query → Parallel 4-signal execution → Per-signal ranking → RRF fusion → Top-K → Graph expansion
The memory layer automatically:
Automatic entity extraction and relationship mapping across your entire development history:
depends-on · implements · tested-by · authored-by · decided-inPowered by code-review-graph, Fulcrum includes a persistent AST-level knowledge graph parsed via Tree-sitter:
Every code change passes through an automated quality pipeline with up to 5 retry iterations:
Code Written → Lint → Type Check → Tests → Coverage Gate → Security Scan
↑ │
└─────────── Automatic Fix & Retry (max 5 iterations) ◄──┘
Scoring engine (scorer.js) — deterministic 0-100 score from 5 components:
| Component | Points | Hard Caps |
|---|---|---|
| Code changes present | 30 | Security violation → max 30 |
| Tests run | 25 | 50%+ test failures → max 40 |
| Test pass rate | 25 | Build broken → max 45 |
| Coverage ≥ 80% | 10 | |
| Clean lint/types | 10 |
Score bands: Poor (<40) · Needs Work (40-59) · Acceptable (60-74) · Good (75-89) · Excellent (≥90)
Additional safeguards:
ECC:SECRET tags before persistenceMulti-agent coordination via claude-flow:
npx claude-flow@alpha swarm start --topology hierarchical # coordinated swarm
npx claude-flow@alpha mcp start # MCP server
npx claude-flow@alpha memory store --key "k" --value "v" # shared vector memory
| Topology | Use Case |
|---|---|
| Hierarchical | Planner → multiple worker agents |
| Mesh | Peer-to-peer collaboration |
| Ring | Sequential pipeline processing |
| Star | Central coordinator with specialists |
3-tier model routing: Tier 1 Agent Booster (WASM, <1ms, $0) → Tier 2 Haiku (~500ms, $0.0002) → Tier 3 Sonnet/Opus (2-5s, complex reasoning)
60+ agent types · HNSW vector memory · Raft/BFT consensus · self-learning loops
Trigger-based automations across the full session lifecycle. All hooks are async with timeouts and continueOnError: true — they never block the user.
| Phase | Count | Purpose |
|---|---|---|
| PreToolUse | 8 | Config protection, privacy gate, security wrappers |
| PostToolUse | 8 | Auto-format, typecheck, quality gate, console.log warnings |
| PreCompact | 2 | Checkpoint + suggest compaction |
| SessionStart | 3 | Session initialization, lifecycle events |
| UserPromptSubmit | 1 | Prompt queueing |
| Stop | 4 | Session save, cost tracking, config guardian |
| SessionEnd | 2 | Session teardown, end markers |
Flag system: minimal · standard · strict profiles via CF_HOOK_PROFILE environment variable.
Sessions persist across restarts with full context recovery:
ecc.session.v1 format with 5 session adapters (canonical, claude-history, dmux-tmux, memory-retrieval, registry)/resume-session restores full context from the last session7 specialized design skills · 67 visual styles · 96 color palettes · 57 font pairings · 25 chart types · 99 UX guidelines · 13 tech stacks
A complete design intelligence suite that transforms any AI coding tool into a full-stack design partner. Every design decision — from color palettes to component tokens to responsive layouts — is backed by searchable databases with BM25-ranked retrieval and stack-specific guidelines.
graph LR
subgraph CORE["🔍 Core Engine"]
UX["ui-ux-pro-max<br/>67 styles · 96 palettes<br/>BM25 search · 13 stacks"]
end
subgraph VISUAL["🎨 Visual Design"]
D["design<br/>55 logo styles<br/>50 CIP deliverables<br/>15 icon styles"]
B["banner-design<br/>22 art directions<br/>multi-format"]
end
subgraph SYSTEM["⚙️ Systems"]
DS["design-system<br/>3-layer tokens<br/>CSS variables"]
ST["ui-styling<br/>shadcn/ui + Tailwind<br/>65+ fonts"]
end
subgraph CONTENT["📝 Content"]
BR["brand<br/>voice · identity<br/>asset management"]
SL["slides<br/>Chart.js · HTML<br/>copywriting"]
end
UX --> D & DS & ST
D --> B
DS --> ST
BR --> SL
| Skill | What It Does | Key Assets |
|---|---|---|
ui-ux-pro-max | Core design intelligence engine — searchable databases with BM25 ranking, design system generation, and priority-based recommendations | 67 styles, 96 palettes, 57 font pairings, 25 chart types, 99 UX guidelines, 13 stack-specific CSV databases, Python BM25 search engine |
design | Comprehensive visual design — logos, corporate identity programs, presentations, banners, icons, social photos | 55 logo styles, 50 CIP deliverables, 15 icon styles, 8 CSV data files, Python generators for logos/CIP/icons |
design-system | Three-layer token architecture (primitive → semantic → component) with CSS variables, spacing scales, and slide generation | 8 slide CSV databases, 9 Node.js/Python scripts for token generation and validation, design token templates |
ui-styling | Beautiful accessible UIs with shadcn/ui (Radix UI + Tailwind), utility-first styling, canvas-based visual designs | 65+ bundled TTF fonts, shadcn/Tailwind reference docs, Python generators for configs |
brand | Brand voice, visual identity, messaging frameworks, and asset management | 11 reference guides (voice framework, color palette management, approval checklist), 4 Node.js scripts, brand guideline templates |
slides | Strategic HTML presentations with data visualization | Chart.js patterns, copywriting formulas, 5 layout strategy guides, responsive slide templates |
banner-design | Multi-format banner design for social media, ads, website heroes, and print | 22 art direction styles, platform-specific dimensions and safe zones |
The design suite includes a built-in Python search engine that provides instant, ranked access to the entire design database:
# Search for design styles matching a query
python scripts/search.py "minimalist dark mode dashboard"
# Generate a complete design system from requirements
python scripts/design_system.py --style modern --palette dark --stack react
The search engine uses BM25 (Best Matching 25) ranking across 32+ CSV databases covering styles, palettes, typography, colors, charts, UX guidelines, and 13 stack-specific recommendation sets (React, Vue, Svelte, Next.js, Flutter, React Native, SwiftUI, HTML+Tailwind, and more).
Every major frontend stack has dedicated design recommendations:
| Stack | Covers |
|---|---|
| React | Component architecture, state patterns, styling approaches |
| Next.js | SSR considerations, image optimization, font loading |
| Vue | Composition API patterns, Vuetify/Quasar integration |
| Svelte | Scoped styling, transitions, SvelteKit layouts |
| Flutter | Material 3, adaptive layouts, platform conventions |
| React Native | NativeWind, platform-specific UX, gesture patterns |
| SwiftUI | SF Symbols, dynamic type, platform idioms |
| HTML + Tailwind | Utility-first, JIT, responsive breakpoints |
| Angular · Nuxt · Astro · Remix · Solid | Framework-specific patterns |
# Install all 7 design skills to your preferred platforms
npx claude-fulcrum skill-install --skills ui-ux-pro-max,design,design-system,ui-styling,brand,slides,banner-design --platform claude,copilot,cursor
# Or install the entire ui-design module (includes all 7)
npx claude-fulcrum install --module ui-design
# Swarm parallel install across all platforms
npx claude-fulcrum skill-swarm --skills ui-ux-pro-max,design,design-system,ui-styling,brand,slides,banner-design --platform all
The design skills automatically activate when working on HTML, CSS, SCSS, React (TSX/JSX), Vue, or Svelte files. Copilot gets dedicated UI/UX instructions with 19 accessibility and design rules that auto-apply based on file type.
Every developer has experienced this:
This is the institutional memory problem. Every codebase has accumulated wisdom — patterns that work, decisions that were made, approaches that failed — but it lives only in people's heads and scattered comments.
Fulcrum's knowledge graph is a persistent institutional memory layer:
auth.ts → decided-in → "Session 2024-01-15: Chose JWT over sessions
because of horizontal scaling requirements"
jwt-service.ts → evolved-from → "token-service.ts (deprecated) —
refactored after CVE-2024-1234"
login-endpoint → tested-by → auth.spec.ts
→ depends-on → jwt-service.ts
→ authored-by → "Claude Code session 2024-02-03"
When you ask "why was this decision made?" — Fulcrum already knows. When a new team member asks "how does auth work?" — the graph traces it. When a bug appears — the evolution history shows what changed.
The codebase remembers. So you don't have to.
26 specialized agents that activate automatically based on context — no manual invocation needed. Each agent is a Markdown file with YAML frontmatter defining its name, description, tools, and model tier (Opus for deep reasoning, Sonnet for most tasks, Haiku for lightweight ops).
Design skills also activate automatically when editing HTML, CSS, SCSS, TSX, JSX, Vue, or Svelte files. See UI/UX Design Intelligence above.
| Agent | Model | Specialty | Activates When |
|---|---|---|---|
planner | Opus | Step-by-step implementation plans with risk analysis | Complex features, multi-file changes |
architect | Opus | System design, scalability, tech selection | Architectural decisions, new systems |
tdd-guide | Sonnet | RED → GREEN → REFACTOR with 80%+ coverage | New features, bug fixes |
code-reviewer | Sonnet | Quality, security, maintainability via git diff | After any code modification |
security-reviewer | Sonnet | OWASP Top 10, secrets, injection, SSRF, crypto | Before commits, auth code |
build-error-resolver | Sonnet | Build/type error resolution with minimal diffs | Any build failure |
e2e-runner | Sonnet | Playwright E2E with screenshots/videos/traces | Critical user flows |
refactor-cleaner | Sonnet | Dead code removal via knip/depcheck/ts-prune | Code maintenance |
doc-updater | Haiku | Documentation and codemap generation | Doc updates |
| Agent | Languages/Frameworks | Key Expertise |
|---|---|---|
python-reviewer | Python, Django, FastAPI, Flask | PEP 8, type hints, Pythonic idioms, security |
go-reviewer | Go, standard library | Idiomatic Go, concurrency, error handling |
kotlin-reviewer | Kotlin, Android, KMP, Compose, Ktor | Coroutine safety, clean architecture |
java-reviewer | Java, Spring Boot, JPA | Layered architecture, concurrency |
rust-reviewer | Rust, async, Tokio | Ownership, lifetimes, unsafe usage |
cpp-reviewer | C++20, CMake | Memory safety, RAII, modern idioms |
| Agent | Fixes |
|---|---|
build-error-resolver | TypeScript/JavaScript build errors |
go-build-resolver | Go compilation, vet, linter issues |
kotlin-build-resolver | Kotlin/Gradle build failures |
java-build-resolver | Java/Maven/Gradle errors |
rust-build-resolver | Cargo build, borrow checker issues |
cpp-build-resolver | CMake, compilation, linker errors |
| Agent | Model | Purpose |
|---|---|---|
database-reviewer | Sonnet | PostgreSQL/Supabase schema & query optimization |
chief-of-staff | Opus | Multi-channel communication triage — email, Slack, LINE, Messenger. 4-tier classification (skip → info → meeting → action) with draft replies. |
loop-operator | Sonnet | Autonomous agent loop monitoring, stall detection & intervention |
harness-optimizer | Sonnet | Harness configuration tuning (cost, reliability, throughput) |
docs-lookup | Sonnet | Live documentation lookup via Context7 MCP (not training data) |
code-graph-reviewer | Sonnet | Knowledge-graph-powered code review — Tree-sitter AST parsing, blast-radius analysis, 14 languages, 6.8x token reduction |
Not sure which agent you need? Fulcrum routes automatically, but here's how it decides:
What are you doing?
│
├── Planning or designing a system?
│ ├── New feature (multi-file) ────────────────→ planner
│ └── Architectural decision ──────────────────→ architect
│
├── Writing code?
│ ├── New feature/bug fix ─────────────────────→ tdd-guide
│ ├── Just finished writing ───────────────────→ code-reviewer (auto)
│ └── Security-sensitive code ─────────────────→ security-reviewer (auto)
│
├── Something broke?
│ ├── TypeScript/JS build error ───────────────→ build-error-resolver
│ ├── Go build error ──────────────────────────→ go-build-resolver
│ ├── Kotlin/Gradle error ─────────────────────→ kotlin-build-resolver
│ ├── Java/Maven error ────────────────────────→ java-build-resolver
│ ├── Rust borrow checker ─────────────────────→ rust-build-resolver
│ └── C++ / CMake error ──────────────────────→ cpp-build-resolver
│
├── Reviewing code?
│ ├── With blast-radius analysis ───────────────→ code-graph-reviewer
│ ├── TypeScript/JavaScript ───────────────────→ code-reviewer
│ ├── Python/Django ───────────────────────────→ python-reviewer
│ ├── Go ──────────────────────────────────────→ go-reviewer
│ ├── Kotlin/Android ──────────────────────────→ kotlin-reviewer
│ ├── Java/Spring Boot ────────────────────────→ java-reviewer
│ ├── Rust ────────────────────────────────────→ rust-reviewer
│ └── C++ ─────────────────────────────────────→ cpp-reviewer
│
├── Working on UI/design?
│ ├── Need a design system ────────────────────→ design-system skill
│ ├── Choosing colors/fonts ───────────────────→ ui-ux-pro-max skill
│ ├── Building components ─────────────────────→ ui-styling skill
│ ├── Brand identity ──────────────────────────→ brand skill
│ └── Presentations ──────────────────────────→ slides skill
│
└── Operations?
├── Database schema/queries ─────────────────→ database-reviewer
├── Communication triage ────────────────────→ chief-of-staff
├── Autonomous loops ────────────────────────→ loop-operator
└── Documentation ───────────────────────────→ doc-updater
Skills are deep domain knowledge that agents draw from — patterns, idioms, testing strategies, design systems, and best practices in Markdown with clear sections (When to Use, How It Works, Examples). They activate contextually — you never need to invoke them manually.
Universal
| Skill | Domain |
|---|---|
coding-standards | Universal TypeScript/React/Node patterns |
api-design | REST naming, status codes, pagination, versioning, rate limits |
frontend-patterns | React, Next.js, state management, performance |
backend-patterns | Architecture, API design, database optimization |
Python
| Skill | Domain |
|---|---|
python-patterns | PEP 8, type hints, Pythonic idioms |
python-testing | pytest, fixtures, mocking, parametrize |
django-patterns | DRF, ORM, signals, middleware |
django-tdd | pytest-django, factory_boy |
django-verification | Migrations, linting, security scans |
Go
| Skill | Domain |
|---|---|
golang-patterns | Idiomatic Go, concurrency, error handling |
golang-testing | Table-driven, subtests, fuzzing, benchmarks |
Kotlin / Android
| Skill | Domain |
|---|---|
kotlin-patterns | Coroutines, null safety, DSL builders |
kotlin-testing | Kotest, MockK, property-based testing |
kotlin-coroutines-flows | Structured concurrency, Flow operators, StateFlow |
kotlin-exposed-patterns | Exposed ORM, HikariCP, Flyway migrations |
kotlin-ktor-patterns | Routing DSL, Koin DI, WebSockets |
compose-multiplatform-patterns | State, navigation, theming |
android-clean-architecture | Modules, UseCases, Repositories |
Java / Spring Boot
| Skill | Domain |
|---|---|
java-coding-standards | Naming, immutability, Optional, streams |
springboot-patterns | Layered services, JPA, async, caching |
springboot-tdd | JUnit 5, Mockito, Testcontainers, JaCoCo |
springboot-verification | Build, static analysis, coverage, security |
Rust
| Skill | Domain |
|---|---|
rust-patterns | Ownership, traits, error handling, async |
rust-testing | Unit, integration, async, property-based |
C++
| Skill | Domain |
|---|---|
cpp-coding-standards | C++ Core Guidelines, modern C++ |
cpp-testing | GoogleTest, CTest, sanitizers |
PHP / Laravel
| Skill | Domain |
|---|---|
laravel-patterns | Eloquent, queues, events, caching |
laravel-tdd | PHPUnit, Pest, factories |
laravel-verification | Env checks, static analysis, coverage |
Perl
| Skill | Domain |
|---|---|
perl-patterns | Modern Perl 5.36+, Moose |
perl-testing | Test2::V0, prove, Devel::Cover |
Other Frameworks
| Skill | Domain |
|---|---|
nextjs-turbopack | Incremental bundling, FS caching |
bun-runtime | Bun vs Node, migration, Vercel |
swiftui-patterns | SwiftUI, actor persistence, concurrency |
| Skill | Domain |
|---|---|
tdd-workflow | Red-Green-Refactor with 80%+ coverage |
e2e-testing | Playwright, Page Object Model, CI/CD |
ai-regression-testing | Sandbox-mode API testing, AI blind spots |
verification-loop | Comprehensive verification system |
eval-harness | Eval-driven development (EDD) framework |
python-testing | pytest, fixtures, mocking, parametrize |
golang-testing | Table-driven, subtests, fuzzing |
kotlin-testing | Kotest, MockK, property-based |
rust-testing | Unit, integration, async, property-based |
cpp-testing | GoogleTest, CTest, sanitizers |
perl-testing | Test2::V0, prove, Devel::Cover |
django-tdd | pytest-django, factory_boy |
laravel-tdd | PHPUnit, Pest, factories |
springboot-tdd | JUnit 5, Mockito, Testcontainers |
django-verification | Migrations, linting, security |
laravel-verification | Static analysis, coverage |
springboot-verification | Build, analysis, coverage |
| Skill | Domain |
|---|---|
claude-api | Messages API, streaming, tool use, vision, batches, prompt caching |
continuous-learning | Auto-extract patterns from sessions |
continuous-learning-v2 | Instinct-based learning with confidence scoring and project scoping |
mcp-server-patterns | Build MCP servers with TypeScript SDK |
dmux-workflows | Multi-agent tmux orchestration |
strategic-compact | Context preservation through task phases |
iterative-retrieval | Progressive context refinement |
deep-research | Multi-source research with firecrawl + Exa MCPs |
market-research | Competitive analysis, market sizing, due diligence |
exa-search | Neural search via Exa MCP |
documentation-lookup | Live docs via Context7 MCP |
fal-ai-media | AI media generation — image, video, audio |
x-api | X/Twitter API integration |
| Skill | Domain |
|---|---|
security-review | OWASP Top 10, auth, input validation, secrets |
plankton-code-quality | Write-time enforcement via hooks |
docker-patterns | Container best practices |
deployment-patterns | CI/CD, blue-green, canary |
database-migrations | Schema migration strategies |
postgres-patterns | PostgreSQL optimization |
django-verification | Django security patterns |
laravel-verification | Laravel security best practices |
springboot-verification | Spring Security, JWT, CORS |
| Skill | Domain |
|---|---|
ui-ux-pro-max | Core design engine — 67 styles, 96 palettes, 57 font pairings, BM25 search, 13 stack databases |
design | Logos (55 styles), corporate identity (50 deliverables), icons (15 styles), banners, social photos |
design-system | Three-layer token architecture, CSS variables, spacing/typography scales, slide generation |
ui-styling | shadcn/ui + Tailwind CSS, 65+ bundled fonts, canvas-based visual designs, dark mode |
brand | Brand voice, visual identity, messaging frameworks, asset management, style guides |
slides | Strategic HTML presentations, Chart.js data visualization, copywriting formulas |
banner-design | Multi-format banners, 22 art direction styles, platform-specific dimensions |
| Skill | Domain |
|---|---|
content-engine | Multi-platform content (X, LinkedIn, TikTok, YouTube) |
article-writing | Long-form content with voice consistency |
market-research | Competitive analysis, market sizing |
investor-materials | Pitch decks, financial models |
investor-outreach | Cold emails, follow-ups |
crosspost | Multi-platform distribution |
video-editing | AI-assisted video workflows (Remotion, ElevenLabs, Descript) |
frontend-slides | Animation-rich HTML presentations |
fal-ai-media | Image, video, audio generation via fal.ai |
x-api | X/Twitter API integration |
Commands are the primary interface — type /command and the right agent activates with the right skills loaded.
/plan — Implementation planning with risk analysis and phases
/multi-plan — Multi-task parallel planning across components
/orchestrate — Agent swarm initialization and coordination
/devfleet — Full development fleet deployment
/projects — Multi-project management and context switching
/tdd — Test-driven development cycle (RED → GREEN → REFACTOR)
/build-fix — Fix build errors automatically with minimal diffs
/code-review — Comprehensive quality + security review via git diff
/code-graph-build — Build/update Tree-sitter knowledge graph (14 languages)
/code-graph-review — Token-efficient delta review with blast-radius analysis
/code-graph-impact — Blast radius impact analysis via knowledge graph
/refactor-clean — Dead code removal (knip, depcheck, ts-prune)
/quality-loop — Iterative quality improvement (max 5 iterations)
/quality-gate — Enforce quality standards before commit
/quality-override — Override quality gates with documented reason
/e2e — E2E test generation & execution with Playwright
/test-coverage — Coverage analysis & gap identification
/verify — Full verification: build + types + lint + tests + secrets + git
/confidence-check — 3-factor confidence scoring before proceeding
/eval — Evaluation-driven development metrics
/python-review — Python PEP 8 + idioms + security review
/go-review — Go idiomatic review + concurrency analysis
/go-build — Fix Go build/vet/linter errors
/go-test — Go test generation (table-driven, subtests)
/kotlin-review — Kotlin/Android/KMP review + coroutine safety
/kotlin-build — Fix Kotlin/Gradle build errors
/kotlin-test — Kotlin test generation (Kotest, MockK)
/rust-review — Rust ownership/lifetime/safety review
/rust-build — Fix Cargo build/borrow checker errors
/rust-test — Rust test generation + property-based testing
/cpp-review — C++ memory safety + modern idioms review
/cpp-build — Fix CMake/compilation/linker errors
/cpp-test — C++ test generation (GoogleTest, CTest)
/gradle-build — Fix Gradle/Android build errors
/save-session — Persist current session state to SQLite
/resume-session — Restore full context from last session
/sessions — List all saved sessions with metadata
/memory-search — 4-signal hybrid search across persistent memory
/learn — Extract reusable patterns from current session
/learn-eval — Evaluate learned patterns against test cases
/checkpoint — Create recovery checkpoint
/skill-create — Generate skills from git history analysis
/skill-health — Monitor skill success rate and relevance
/evolve — Evolve skills and agents based on usage data
/promote — Promote instinct to permanent skill
/instinct-status — View instinct confidence scores and maturity
/instinct-export — Export instincts for sharing across projects
/instinct-import — Import instincts from another project
/loop-start — Start autonomous agent loop with safety bounds
/loop-status — Monitor loop progress, detect stalls
/multi-execute — Parallel task execution across agents
/multi-backend — Backend multi-agent coordination
/multi-frontend — Frontend multi-agent coordination
/multi-workflow — Full-stack workflow orchestration
/harness-audit — Audit harness installation and health
/model-route — Route task to optimal model (Haiku/Sonnet/Opus)
/pm2 — PM2 process management integration
/docs — Fetch live documentation via Context7 MCP
/update-docs — Sync documentation with codebase changes
/update-codemaps — Regenerate architecture codemaps
/prompt-optimize — Optimize prompts for better AI responses
/setup-pm — Configure package manager (npm/pnpm/yarn/bun)
/aside — Start sidebar conversation without context pollution
/attempt — Try approach without full planning commitment
/claw — NanoClaw terminal REPL with pattern learning
Coding standards enforced automatically across every AI tool. Each ruleset includes 5 files: coding-style · testing · security · patterns · hooks. Rules auto-apply based on file extension.
| Rule Set | Key Enforcements |
|---|---|
| Common | Immutability, error handling, input validation, 80%+ coverage, security checklist |
| TypeScript | Strict mode, ESM imports, React/Next.js patterns, Zod validation, Jest/Vitest |
| Python | PEP 8, type hints, dataclasses, Pythonic idioms, pytest, flake8/mypy |
| Go | Standard library first, explicit error handling, table-driven tests, go fmt/vet |
| Kotlin | Coroutines, null safety, sealed classes, Compose state, ktlint/detekt |
| C++ | Modern C++20, RAII, smart pointers, no raw new/delete, clang-format/tidy |
| Perl | Modern Perl 5.36+, strict/warnings, Moose/Moo, perltidy/perlcritic |
| PHP | PSR-12, Laravel conventions, Composer autoloading |
| Swift | Swift 6.2 concurrency, protocol-oriented, actor isolation |
Each rule set covers: coding style, testing requirements, security practices, hook integrations, and common patterns.
| Operation | Time | Notes |
|---|---|---|
| Memory search (4-signal) | <50ms | FTS5 + vector + graph parallel |
| Session restore | <200ms | Full context from last session |
| Quality loop (pass) | <30s | Average first-pass on clean code |
| Agent activation | <100ms | Context-based auto-selection |
| Embedding generation | <3ms | ONNX local, no API call |
| Hook execution (full cycle) | <500ms | 33 hooks, all async |
| Skill install (all platforms) | <10s | Sequential mode |
| Skill install (swarm mode) | <3s | 6 parallel workers |
Benchmarked on MacBook M3, Node.js 20, 10k observation database
| Capability | ECC | Claude Fulcrum |
|---|---|---|
| Agents | 21 | 26 |
| Skills | ~80 | 119 (includes 7 UI/UX design intelligence skills) |
| Commands | ~40 | 66 |
| Platforms | 4 | 6 (Claude Code, Codex, Cursor, Copilot, OpenCode, Antigravity) |
| Design Intelligence | None | 7 skills — 67 styles, 96 palettes, 57 font pairings, BM25 search |
| Language Rules | 5 | 9 (each with 5 files: style, testing, security, patterns, hooks) |
| Memory | Per-platform | Unified hybrid (FTS5 + vector + knowledge graph + 50+ API methods) |
| Search | Keyword | 4-signal neural (lexical + recency + structure + vector via RRF) |
| Embeddings | None | 384-dim ONNX (Xenova/all-MiniLM-L6-v2) + hash fallback |
| Orchestration | Commands only | Full swarm (claude-flow, 4 topologies, 3-tier model routing) |
| Quality Loop | Partial | Automated (scorer + confidence gate + cross-model auditor + policy validators) |
| Knowledge Graph | No | Yes (12 entity types, 5 relationship types, multi-hop reasoning) |
| Institutional Memory | No | Yes — decisions, evolution history, author tracing |
| Hooks | Few | 33 hooks across 7 lifecycle phases with flag-based profiles |
| Tests | Some | 1,536 passing across 76 test files |
| Schemas | None | 13 JSON Schemas validated via Ajv at runtime |
| CI Validators | None | 8 validators (agents, skills, commands, rules, hooks, manifests, paths) |
| Cross-platform Guide | No | Yes (full integration docs + extreme dev playbook) |
The four directions dominating AI tooling in 2026:
| Direction | What It Means | How Fulcrum Addresses It |
|---|---|---|
| Agentic execution | AI that acts autonomously, not just suggests | 25 auto-activating agents, quality loop, autonomous loops |
| Workflow orchestration | Coordinating multiple AI tools and models | 6-platform unified layer, swarm coordination, 3-tier routing |
| Data and context | AI that remembers and learns from your history | 4-signal hybrid memory, knowledge graph, institutional memory |
| Multimodal generation | Code + design + content in one flow | 7 design skills, BM25 search, 96 palettes, 57 font pairings |
Fulcrum is the only developer harness that addresses all four simultaneously.
| Platform | Config Location | Components | Best For |
|---|---|---|---|
| Claude Code | ~/.claude/ | Full 26 agents, 122 skills, 9 rulesets, 33 hooks | Deep workflows, TDD, planning, security |
| Codex CLI | ~/.codex/ | Core agents + rules | Fast parallel execution, batch tasks |
| Cursor | .cursor/ | 11 agents, full rules | Visual IDE, real-time coding |
| GitHub Copilot | .github/ | 11 agents, 30 prompts, 8 language instructions | Inline completions, chat |
| OpenCode | .opencode/ | Core agents + skills | Open-source, extensible |
| Antigravity | .agent/ | Agents, skills, flattened rules | Workflow-native agent IDE |
Shared context: All platforms read from and write to the same memory layer. Patterns learned in Claude Code are available in Copilot completions. Plans created in one tool execute in another.
Deploy skills to any platform individually or in bulk using the skill installer CLI:
# Sequential — installs one skill at a time per platform
cf-skill-install --all --platform all
cf-skill-install --skills tdd-workflow,security-review --platform claude,copilot
# Swarm mode — one worker thread per platform, all run in parallel
cf-skill-swarm --all --platform all
cf-skill-swarm --skills api-design --platform claude,cursor,codex
# Inspect available skills and platforms
cf-skill-install --list
cf-skill-install --list-platforms
cf-skill-install --dry-run --all --platform all --json
Swarm architecture: Coordinator spawns one worker_threads worker per platform. Each worker deploys all requested skills independently. Results are aggregated when all workers complete.
| Guide | Description |
|---|---|
| Complete Architecture | Full technical reference for every subsystem — memory, quality, hooks, orchestration, install |
| Cross-Platform Integration | How all 6 platforms share one orchestration layer |
| Extreme Dev Playbook | Daily workflow for all platforms working together |
| Shortform Guide | Setup, foundations, philosophy |
| Longform Guide | Token optimization, memory persistence, evals |
| Security Guide | Scanning, secret management, OWASP patterns |
| Copilot Integration | GitHub Copilot agent, prompt, and instruction setup |
| Selective Install Design | Profile-based installation (core, developer, security, research, full) |
| Skill Installer CLI | Install any/all skills to any/all platforms (sequential + swarm) |
| Session Adapter Contract | Canonical session schema and adapter protocol |
| Troubleshooting | Common issues and fixes |
claude-fulcrum/
├── agents/ # 26 specialized agents (YAML frontmatter + Markdown)
├── skills/ # 119 workflow skills (domain knowledge modules)
│ ├── ui-ux-pro-max/ # Design engine: 67 styles, BM25 search, 13 stack databases
│ ├── design/ # Logos, CIP, icons, banners, social photos
│ ├── design-system/ # Token architecture, CSS variables, slide generation
│ ├── ui-styling/ # shadcn/ui + Tailwind, 65+ fonts, canvas designs
│ ├── brand/ # Voice, identity, messaging, asset management
│ ├── slides/ # HTML presentations, Chart.js, copywriting
│ ├── banner-design/ # Multi-format banners, 22 art direction styles
│ └── ... # 112 more: languages, testing, security, AI, business
├── commands/ # 66 slash commands
├── hooks/ # hooks.json + 33 hook scripts across 7 lifecycle phases
├── rules/ # 9 language rule sets (5 files each)
│ ├── common/ # Universal: immutability, testing, security, patterns, hooks
│ ├── typescript/ # TypeScript/React/Next.js (strict, ESM, Zod)
│ ├── python/ # Python/Django (PEP 8, type hints, pytest)
│ ├── golang/ # Go (stdlib-first, table-driven tests, go fmt/vet)
│ ├── kotlin/ # Kotlin/Android/KMP (coroutines, Compose, ktlint)
│ ├── cpp/ # C++20 (RAII, smart pointers, clang-format)
│ ├── perl/ # Perl 5.36+ (strict/warnings, Test2)
│ ├── php/ # PHP/Laravel (PSR-12, Composer)
│ └── swift/ # Swift 6.2 (concurrency, actors, SwiftLint)
├── scripts/ # 144 scripts
│ ├── hooks/ # 33 hook scripts (session, quality, security, formatting)
│ ├── lib/ # Core libraries
│ │ ├── state-store/ # SQLite state store (sql.js WASM, 4 migrations)
│ │ ├── memory/ # 7 memory modules (search, graph, dedup, evolution)
│ │ ├── quality/ # 8 quality modules (scorer, confidence, auditor, policies)
│ │ ├── skill-evolution/ # 6 modules (dashboard, health, tracker, versioning)
│ │ └── session-adapters/ # 5 adapters (canonical, history, dmux, memory, registry)
│ ├── ci/ # 8 CI validators (agents, skills, commands, rules, hooks)
│ └── utils/ # Embeddings (ONNX 384-dim), package manager, validators
├── schemas/ # 13 JSON Schemas (Ajv validation at runtime)
├── tests/ # 1,536 tests across 76 files (80%+ coverage)
├── orchestration/ # claude-flow swarm orchestration (4 topologies)
├── mcp-configs/ # MCP server configurations (8 servers)
├── config/ # Hook control plane (protected paths, policy bundles)
├── docs/ # Comprehensive guides + ARCHITECTURE.md
│ ├── ARCHITECTURE.md # Full technical reference (16 sections)
│ └── ... # Cross-platform, security, playbooks, i18n (ja, ko, zh)
├── .github/ # GitHub + Copilot config (11 agents, 30 prompts, 8 instructions)
├── .codex/ # Codex CLI config
├── .cursor/ # Cursor config
├── .opencode/ # OpenCode config
├── install.sh # Unix installer (language-specific)
├── install.ps1 # Windows installer
├── CLAUDE.md # Claude Code instructions
├── AGENTS.md # Agent routing and orchestration rules
└── package.json # Project manifest
Configured in mcp-configs/:
| Server | Purpose |
|---|---|
| claude-flow | Multi-agent swarm orchestration (4 topologies, HNSW memory) |
| memory | Persistent memory across sessions (SQLite + FTS5 + vector) |
| sequential-thinking | Step-by-step reasoning chains |
| context7 | Live library documentation lookup |
| playwright | Browser automation & E2E testing |
| firecrawl | Web scraping & research |
| supabase | Database operations |
| github | Repository management |
CLAUDE_PACKAGE_MANAGER=pnpm # Override package manager (npm, pnpm, yarn, bun)
CF_HOOK_PROFILE=standard # Hook profile (minimal, standard, strict)
CF_DISABLED_HOOKS=cost-tracker # Comma-separated hook names to disable
Protected paths, policy bundles, and context budget management in config/hook-control-plane.json:
npm test # Run all 1,536 tests
npm run coverage # Coverage report (80%+ required)
npm run lint # Lint everything
npm run harness:audit # Audit harness configuration
| Category | Files | Tests Cover |
|---|---|---|
| Core libraries | 26 | Install, session, state store, quality, package manager |
| Hook handlers | 16 | Lifecycle, quality gates, formatting, config protection |
| CLI scripts | 17 | Installation, repair, diagnostics, status |
| Memory system | 7 | Graph retrieval, dedup, evolution, hybrid search, time-aware traversal |
| Quality modules | 5 | Confidence scoring, model auditing, policy validation |
| Agent system | 4 | Agent selector, handoff contracts, persona router |
| CI validators | 1 | Structural integrity checks |
| Integration | 1 | Full pipeline end-to-end |
| Total | 76 files | 1,536 passing |
Run on every commit to enforce structural integrity:
┌─────────────┐
│ Query │
└──────┬──────┘
│
┌────────────┼────────────┐
│ │ │
┌─────▼────┐ ┌────▼─────┐ ┌───▼────────┐
│ FTS5 / │ │ Vector │ │ Knowledge │
│ LIKE │ │ Cosine │ │ Graph Walk │
│ Search │ │ Sim 384d │ │ BFS + DFS │
└─────┬────┘ └────┬─────┘ └───┬────────┘
│ │ │
┌─────▼────┐ ┌────▼─────┐ ┌───▼────────┐
│ Lexical │ │ Semantic │ │ Structure │
│ Ranking │ │ Ranking │ │ + Recency │
└─────┬────┘ └────┬─────┘ └───┬────────┘
│ │ │
└────────────┼────────────┘
│
┌────────────▼────────────┐
│ Reciprocal Rank Fusion │
│ (RRF merge + rerank) │
└────────────┬────────────┘
│
┌────────────▼────────────┐
│ Top-K Results │
│ + Graph Expansion │
└─────────────────────────┘
Scoring formula: Each signal produces a ranked list. RRF merges them: score(d) = Σ 1/(k + rank_i(d)) where k=60 (standard RRF constant). The fused score combines all four signals without requiring calibration.
The deterministic scorer computes a 0-100 score:
Base Score = (code_changes × 30) + (tests_run × 25) + (pass_rate × 25) + (coverage × 10) + (clean × 10)
Hard Caps Applied:
if security_violation → score = min(score, 30)
if test_failure_rate > 50% → score = min(score, 40)
if build_broken → score = min(score, 45)
Trend Detection:
isImproving = score[n] > score[n-1] > score[n-2]
isOscillating = score[n] > score[n-1] AND score[n-1] < score[n-2]
isStagnating = |score[n] - score[n-1]| < 2 for 3+ iterations
Termination: Loop ends on QUALITY_MET (score ≥ threshold) · OSCILLATION · STAGNATION · MAX_ITERATIONS (5) · ERROR
Profile Selection (core | developer | security | research | full)
│
▼
Module Resolution (DFS topological sort)
│
▼
Target Adapter Selection (claude-home | cursor-project | codex-home | opencode-home | antigravity-project)
│
▼
Operation Strategy (preserve-relative-path | flatten-copy | sync-root-children)
│
▼
File Copy with Conflict Resolution
Profiles: Core (6 modules, essentials) → Developer (9 modules, +testing +docs) → Security (7 modules, +scanning) → Research (9 modules, +research skills) → Full (19 modules, everything).
SessionStart hooks fire (3)
→ session-lifecycle.js creates session record
→ context loaded from last session
→ memory notes retrieved
For each user interaction:
UserPromptSubmit hook (1)
→ prompt queued for processing
For each tool call:
PreToolUse hooks fire (8)
→ privacy-gate strips secrets
→ protect-configs blocks protected file edits
→ repo-hook-trust validates hook integrity
Tool executes (read, write, edit, bash, grep)
PostToolUse hooks fire (8)
→ post-edit-format auto-formats code
→ post-edit-typecheck runs type checker
→ quality-gate evaluates changes
→ check-console-log warns on debug statements
Stop hooks fire (4)
→ session state saved
→ cost-tracker logs metrics
→ config-guardian validates config integrity
SessionEnd hooks fire (2)
→ session-end-marker records termination
→ cleanup
Star this repo to follow progress. Open an issue to influence the roadmap.
Claude Fulcrum extends and integrates:
The unified memory layer, 4-signal hybrid search, knowledge graph, quality enforcement system, 33-hook lifecycle, Copilot shared memory integration, institutional memory concept, UI/UX design intelligence suite, and code-review-graph platform integration are original work in this repo.
See CONTRIBUTING.md for details.
# Fork → branch → implement → test → PR
git checkout -b feat/my-feature
npm test # all 1,536 tests must pass
File formats:
| Type | Format |
|---|---|
| Agents | Markdown with YAML frontmatter (name, description, tools, model) |
| Skills | Markdown with sections (When to Use, How It Works, Examples) |
| Commands | Markdown with description frontmatter |
| Rules | Markdown in rules/<language>/ |
Naming convention: lowercase with hyphens — python-reviewer.md, tdd-workflow.md
claude-code · claude-code-agents · ai-dev-workflow · claude-code-setup · codex-cli · cursor-ai · github-copilot · opencode · claude-flow · ai-agents · mcp-server · developer-tools · tdd · code-review · multi-agent · agent-orchestration · vector-search · knowledge-graph · swarm-intelligence · design-system · ui-ux
| Metric | Count |
|---|---|
| Specialized Agents | 26 |
| Workflow Skills | 122 |
| Slash Commands | 66 |
| Language Rule Sets | 9 |
| Rule Files (total) | ~50 |
| Supported Platforms | 6 |
| Lifecycle Hooks | 33 |
| Hook Phases | 7 |
| Utility Scripts | 144 |
| Quality Modules | 8 |
| Memory Modules | 7 |
| CI Validators | 8 |
| JSON Schemas | 13 |
| Test Cases | 1,536 |
| Test Files | 76 |
| Search Signals | 4 |
| Entity Types | 12 |
| Embedding Dimensions | 384 |
| Design Styles | 67 |
| Color Palettes | 96 |
| Font Pairings | 57 |
| UX Guidelines | 99 |
| Tech Stack Databases | 13 |
| Copilot Agents | 12 |
| Copilot Prompts | 34 |
| Install Profiles | 5 |
| MCP Servers | 9 |
MIT — use it, fork it, ship it.
Claude Fulcrum — where every AI coding tool becomes part of the same team — from architecture to pixel-perfect design.
The codebase remembers. So you don't have to.
FAQs
The operating system for AI-powered development — 25 agents, 112 skills, 62 commands, 4-signal hybrid memory, knowledge graphs, and cross-platform support for Claude Code, Codex, Cursor, Copilot & OpenCode
We found that claude-fulcrum demonstrated a healthy version release cadence and project activity because the last version was released less than a year ago. It has 1 open source maintainer collaborating on the project.
Company News
Socket is joining the OpenJS Security Stewardship Program to fund Node.js vulnerability research, maintainer remediation, and security releases.
Security News
Two compromised GitHub Actions were re-enabled with malicious tags intact, exposing thousands of downstream repositories to Mini Shai-Hulud.
Research
/Security News
A malicious Firefox extension fetches its payload after installation to evade detection, steal Google session cookies, and automate account takeover.