The Agent Skills engine — extracted, standalone, and universal.
Claude Code, GitHub Copilot, OpenAI Codex, and Cursor all have a skill engine built in — the agent discovers SKILL.md playbooks, reads them on demand, and follows structured instructions to produce consistent, high-quality output. That's why they can generate professional documents, presentations, and code without being told how every time.
But the skill engine is locked inside each product.
If you're building with LangChain, CrewAI, AutoGen, Smolagents, or any custom agent — you get nothing. No skill discovery, no progressive disclosure, no structured playbooks. You'd have to build the entire engine from scratch.
FastSkills extracts the skill engine into a standalone MCP server. One command gives any agent the same capability:
uvx fastskills --skills-dir ./skillsThat's it. Your agent can now discover, read, and execute any SKILL.md — the same open standard used by Claude Code, Copilot, Codex, OpenClaw, and nanobot.
No framework adoption. No code changes. No SDK integration. Just MCP.
WITHOUT FastSkills:
Claude Code ──── has skill engine ──── can use 40,000+ skills
Copilot ──── has skill engine ──── can use 40,000+ skills
Codex CLI ──── has skill engine ──── can use 40,000+ skills
OpenClaw ──── has skill engine ──── can use 40,000+ skills
nanobot ──── has skill engine ──── can use 40,000+ skills
─────────────────────────────────────────────────────────────
LangChain ──── nothing ──── ✗
CrewAI ──── nothing ──── ✗
AutoGen ──── nothing ──── ✗
Your agent ──── nothing ──── ✗
WITH FastSkills:
Any agent ──── FastSkills (MCP) ──── can use 40,000+ skills ✓
You can — and they're great. But they solve a different problem:
| OpenClaw | nanobot | FastSkills | |
|---|---|---|---|
| What it is | Full AI assistant platform | Lightweight AI assistant | Standalone skill engine |
| To use skills | Adopt the entire platform (Gateway, config, workspace) | Install the framework, use its agent loop | One command or one JSON block |
| Lock-in | High — skills require OpenClaw runtime | Medium — skills require nanobot agent | None — standard MCP protocol |
| Works with your agent? | No — you use OpenClaw's agent | No — you use nanobot's agent | Yes — any MCP client |
| Lines of code | 430,000+ | ~4,000 | ~800 |
| Install | npm install -g openclaw && openclaw onboard --install-daemon |
pip install nanobot && nanobot onboard |
uvx fastskills --skills-dir ./skills |
OpenClaw and nanobot are platforms — you adopt their agent, their runtime, their ecosystem. FastSkills is a building block — it gives your existing agent skill capabilities without replacing anything.
Add one block to your MCP client config (Claude Desktop, Cursor, VS Code, Windsurf, Goose, or any custom agent):
{
"mcpServers": {
"fastskills": {
"command": "uvx",
"args": ["fastskills", "--skills-dir", "/path/to/skills", "--workdir", "/path/to/output"]
}
}
}Your agent now has skill discovery, skill reading, and a full execution toolkit (bash, file creation, string replacement) — matching the tool surface that Claude Code provides natively.
What's
uvx? It's a tool from uv that runs Python packages in isolated environments — no install step needed. Install it withcurl -LsSf https://astral.sh/uv/install.sh | shorbrew install uv.
# Clone the repo
git clone https://github.com/nj19257/FastSkills.git
cd FastSkills
# Start chatting (requires uv: https://docs.astral.sh/uv/)
uv run fastskills_cliOn first launch you'll be prompted for your OpenRouter API key and model. The bundled 18 skills are automatically discovered.
| Command | What it does |
|---|---|
/help |
Show all commands |
/skills |
List available skills |
/search <query> |
Search cloud skill catalog |
/model or /settings |
Change AI model and settings |
/sessions |
List saved sessions |
/load <N> |
Load a saved session |
/save |
Force-save current session |
/clear |
Clear chat and start new session |
/status |
Show connection info |
Ctrl+H |
Toggle sidebar |
The same progressive disclosure pattern used by Claude Code, Copilot, and Codex:
┌──────┐ ┌──────┐ ┌──────┐ ┌────────┐ ┌─────────┐
│ Scan │───►│Match │───►│ Read │───►│ Follow │───►│ Deliver │
└──────┘ └──────┘ └──────┘ └────────┘ └─────────┘
metadata best fit SKILL.md run scripts output
~100 tok agent full body bash/create consistent
per skill decides on demand on demand quality
- Scan —
list_skillsreturns names + descriptions (~100 tokens per skill) - Match — The agent decides which skill fits the task
- Read —
view(path=...)reads the fullSKILL.mdinto context - Follow —
bash_toolruns scripts,file_createwrites output - Deliver — Consistent, high-quality results every time
The agent only loads what it needs, when it needs it. No context window bloat.
FastSkills includes a gold-standard system prompt at prompt/gold_standard_prompt.yaml that teaches any agent the discover → read → execute workflow. Use it directly or reference it to build your own.
┌─────────────────┐ MCP ┌──────────────┐ filesystem ┌──────────────┐
│ Your Agent │◄────────────────►│ FastSkills │◄──────────────────►│ skills/ │
│ (any MCP │ protocol │ MCP Server │ read SKILL.md │ ├── pptx/ │
│ client) │ │ (FastMCP) │ run scripts │ ├── docx/ │
└─────────────────┘ └──────────────┘ │ ├── pdf/ │
│ └── ... │
└──────────────┘
| Tool | Purpose |
|---|---|
list_skills |
Discover available skills with name, description, and SKILL.md path |
view |
Read files (including SKILL.md), list directories, inspect images |
bash_tool |
Execute shell commands and skill scripts (120s timeout) |
file_create |
Create output files with content |
str_replace |
Make targeted string replacements in existing files |
| Tool | Purpose |
|---|---|
search_cloud_skills |
Search the SkillsMP cloud catalog by keyword |
install_cloud_skill |
Install a skill from the cloud catalog into the local skills directory |
To enable cloud tools, pass SKILLSMP_API_KEY via the MCP client config:
{
"mcpServers": {
"fastskills": {
"command": "uvx",
"args": ["fastskills", "--skills-dir", "/path/to/skills", "--workdir", "/path/to/output"],
"env": {
"SKILLSMP_API_KEY": "sk_live_your_key_here"
}
}
}
}| Category | Skills |
|---|---|
| Documents | pptx · docx · pdf · xlsx |
| Design | theme-factory · brand-guidelines · canvas-design · frontend-design · frontend-slides · algorithmic-art |
| Authoring | doc-coauthoring · internal-comms · web-artifacts-builder · slack-gif-creator |
| Developer | mcp-builder · skill-creator · webapp-testing · duckduckgo-websearch |
Drop a folder with a SKILL.md into your skills directory. FastSkills picks it up automatically:
~/.fastskills/skills/
├── pptx/
│ └── SKILL.md
├── my-custom-skill/
│ ├── SKILL.md
│ ├── scripts/
│ │ └── generate.py
│ └── references/
│ └── style-guide.md
└── ...
Example SKILL.md
---
name: api-documentation
description: Generate consistent API documentation following team standards.
Use when writing docs for REST endpoints, SDKs, or internal APIs.
---
# API Documentation Skill
## When to Use
Use this skill when the user asks to document an API, generate endpoint
references, or create SDK documentation.
## Instructions
1. Read the source code or endpoint definitions
2. Extract parameters, return types, and error codes
3. Generate documentation following the template in ./references/template.md
4. Include code examples for each endpointThe MCP server loads skills from a single directory specified by --skills-dir. The TUI (fastskills_cli) auto-resolves from two locations:
| Location | Used by | Description |
|---|---|---|
--skills-dir path |
MCP server | The single source of skills for the MCP server (required flag) |
Repo-bundled skills/ |
TUI | 18 bundled skills shipped with FastSkills |
~/.fastskills/skills/ |
TUI | User-global skills auto-discovered by the TUI |
You can also use skills from Anthropic's skills repo, SkillsMP (cloud catalog), or any community source.
FastSkills uses the same open standard adopted across the industry:
| Platform | Skills Support | How |
|---|---|---|
| Claude Code | ✅ Native | .claude/skills/ directory |
| GitHub Copilot | ✅ Native | .github/skills/ directory |
| OpenAI Codex CLI | ✅ Native | Same SKILL.md format |
| Cursor | ✅ Native | Built-in skill engine |
| OpenClaw | ✅ Native | AgentSkills-compatible folders |
| nanobot | ✅ Native | Bundled + custom skills |
| LangChain / CrewAI / AutoGen | ❌ → ✅ | Via FastSkills |
| Any MCP-compatible agent | ❌ → ✅ | Via FastSkills |
Skills are portable. Write once, use everywhere.
| Flag | Description | Default |
|---|---|---|
--skills-dir |
Root directory containing skill folders | (required) |
--workdir |
Working directory for command execution and file output | Current directory |
# Start the MCP server
uvx fastskills --skills-dir /path/to/skills
# With a working directory for file output and command execution
uvx fastskills --skills-dir /path/to/skills --workdir /path/to/outputTip: Pair FastSkills with a web search MCP server (like mcp-server-fetch) so your agent can research topics alongside executing skills. Skills handle the "how," web search handles the "what."
Contributions welcome — new skills, core improvements, or docs:
- Fork the repository
- Create a feature branch (
git checkout -b feature/amazing-skill) - Commit your changes
- Open a Pull Request
MIT License — see LICENSE for details.
- Follows the Agent Skills open standard
- Built with FastMCP
- Inspired by the skill systems in Claude Code, OpenClaw, and nanobot
The skill engine is locked inside coding agents. FastSkills sets it free.
⭐ Star this repo if FastSkills is useful to you