SkillFlow

Skills flowing between your agents.
The universal skill manager for the multi-agent era.

How it works • Supported agents • Quick start • Packages • Roadmap

---

What is SkillFlow?

SkillFlow is a unified tool to install, sync, audit, and manage Agent Skills across every AI coding agent — from a single source of truth.

Today, skills live in scattered directories, break through symlinks, and have no visibility across agents. SkillFlow fixes this by providing an MCP server for real-time agent integration, an interactive CLI for terminal workflows, and a React dashboard for visual management.

The problem: You install 100+ skills via npx @anthropic-ai/find-skills, but Cowork can't see symlinks, Codex uses a different directory, and Cursor ignores both. You have zero visibility into what's installed, what's duplicated, and what's broken.

The solution: SkillFlow syncs real copies to every agent's skill directory, validates SKILL.md against the open spec, detects conflicts, and gives you one place to manage everything.

How does SkillFlow work?

SkillFlow operates in three layers, each designed for minimal context overhead (following the progressive disclosure pattern recommended by the Agent Skills spec):

┌─────────────────────────────────────────────────────┐
│                  @skillflow/dashboard                │
│           React UI · Visual management               │
├─────────────────────────────────────────────────────┤
│                   @skillflow/mcp                     │
│     MCP Server · Real-time agent integration         │
├─────────────────────────────────────────────────────┤
│                   @skillflow/core                    │
│  Discovery · Validation · Sync · Conflict detection  │
├─────────────────────────────────────────────────────┤
│                   @skillflow/cli                     │
│      Interactive terminal · Low-token interface      │
└─────────────────────────────────────────────────────┘

Why this architecture? Traditional MCP servers dump 44,000–150,000 tokens of JSON schemas into the context window before any work begins. SkillFlow's CLI-first approach keeps agent interactions under 1,300 tokens per command, preserving your model's reasoning capacity. The MCP layer is available when you need structured auth and audit trails for enterprise environments.

What agents does SkillFlow support?

Agent	Skill Directory	Status
Claude Code	~/.claude/skills/	Supported
Cowork	Via plugins	Supported
OpenAI Codex CLI	~/.codex/skills/	Supported
Google Antigravity	~/.gemini/antigravity/skills/	Supported
Cursor	~/.cursor/skills/	Supported
Windsurf	~/.codeium/windsurf/skills/	Supported
GitHub Copilot	.github/skills/	Supported
VS Code Copilot	.vscode/skills/	Planned
Warp	~/.warp/skills/	Planned

SkillFlow follows the Agent Skills Specification — the open standard adopted by Anthropic, Microsoft, OpenAI, Google, Cursor, Atlassian, Figma, and 40+ agents.

Quick start

# Install globally
npm install -g skillflow

# See what you have
skillflow scan

# Sync skills to all agents (real copies, no symlinks)
skillflow sync --all

# Audit for broken, duplicated, or invalid skills
skillflow audit

# Launch the dashboard
skillflow dashboard

As an MCP server

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "skillflow": {
      "command": "npx",
      "args": ["-y", "@skillflow/mcp"]
    }
  }
}

As an Agent Skill

skillflow install-self
# Installs a lightweight skill that teaches agents to use the CLI
# ~200 tokens loaded at startup — full docs on demand via --help

Packages

Package	Description	npm
skillflow	Meta-package (installs CLI + Core)
@skillflow/core	Engine — discovery, validation, sync, conflict detection
@skillflow/cli	Interactive terminal interface
@skillflow/mcp	MCP server for structured agent integration
@skillflow/dashboard	React dashboard for visual management

How is SkillFlow different from existing tools?

Feature	SkillFlow	npx skills (Vercel)	ASM	SkillPort
Multi-agent sync	All 7+ agents	Claude only	15 agents (audit only)	Claude + Codex
Real copies (no symlinks)	Yes	No (symlinks break)	N/A	Partial
MCP server	Yes	No	No	Yes
Interactive CLI	Yes	No	Python CLI	No
Visual dashboard	Yes	No	No	No
Skill validation	Full spec	Basic	Audit only	Basic
Context-efficient	~200 tokens	N/A	N/A	Heavy JSON

Roadmap

• Reserve npm packages and GitHub repo
• v0.1 — Core engine: scan, validate, sync across 7 agents
• v0.2 — CLI: interactive terminal with skillflow scan, sync, audit
• v0.3 — MCP server: agent-callable tools with progressive disclosure
• v0.4 — Dashboard: React UI with real-time skill health monitoring
• v0.5 — Registry integration: search and install from skills.sh
• v1.0 — Stable release with full B2A optimization and llms.txt

Architecture decisions

SkillFlow is built on three key insights from the 2026 agent ecosystem:

1. CLI-first, MCP-optional — Agents natively excel at terminal commands. A skillflow sync costs ~200 tokens vs. 44,000+ for loading a full MCP schema. We use the MCP layer only when enterprises need OAuth, audit trails, and multi-tenant isolation.

2. Real copies over symlinks — Symlinks break across VM boundaries (Cowork), Docker containers, and remote environments. SkillFlow always writes real files to each agent's directory.

3. Progressive disclosure — Following the Agent Skills spec, only name + description are loaded at startup (~100 tokens). The full SKILL.md loads on activation. Supporting files load on demand.

Tech stack

TypeScript · MCP SDK · Zod · Commander.js · React · Vite · Agent Skills Spec

Contributing

SkillFlow is open source and welcomes contributions. See CONTRIBUTING.md for guidelines.

git clone https://github.com/MAUGUS2/skillflow.git
cd skillflow
npm install
npm run dev

License

MIT — use it, fork it, make it better.

---

_{Built with care by Fluity Agents
Part of the Fluity ecosystem — ShellGate · SkillFlow · more coming soon}