Skills MCP Server
Centralized, sandboxed MCP server to manage, execute, and expose advanced Agent Skills to any MCP-capable framework or low-code platform.
README
๐ ๏ธ Skills MCP Server
A centralized, sandboxed Model Context Protocol (MCP) server designed to manage, execute, and expose advanced Agent Skills to any MCP-capable framework or low-code platform (like n8n, custom agents, etc.).
๐ก The Problem It Solves
Currently, many popular orchestration frameworks and low-code platforms only support basic tools or MCP. However, modern AI agents require fully encapsulated, complex agent skills (prompts, tool chains, and logic combined). Agno framework natively supports these advanced agent skills, but getting them into platforms like n8n or custom agentic systems has historically been a challenge.
skills_mcp_server is the bridge. By wrapping Agno skills in an isolated FastAPI server and exposing them via the universal Model Context Protocol (MCP), any system that supports MCP can now instantly leverage complex agent skills.
โจ Key Advantages
- ๐ฏ Selective Skill Exposure (Context Optimization): Install a massive library of skills on the server, but expose only the exact ones a specific agent needs. By tailoring the skill list per agent, you prevent LLM context bloat, save tokens, and drastically improve agent accuracy.
- ๐๏ธ Centralized Skill Hub: Stop scattering custom scripts and functions across different repositories or n8n nodes. Manage your entire organization's AI capabilities in one unified, easily updatable server.
- ๐ Sandboxed Execution: All skills run in an isolated environment. You can safely install third-party community skills without risking the integrity of your host application or primary infrastructure.
- ๐ฆ Automatic Dependency Isolation: Skills that include a
requirements.txtfile are automatically provisioned with a dedicated virtual environment usinguv. This ensures perfect isolation and clean uninstalls without bloating the main server's dependencies. - ๐ง Powered by Agno: Go beyond simple tools. Leverage Agno to build rich, stateful agent skills that can handle complex multi-step reasoning before returning the final payload via MCP.
- ๐ฆ Dynamic Hot-Loading: Add, update, or remove skills on the fly using
.zipfiles, custom.skillpackages, or direct download URLs without ever restarting the server.
๐๏ธ Architecture Concept
The system operates on a dual-interface architecture:
- REST API (The Manager): A control plane to dynamically install, update, list, and delete skills. It acts as your private skill store.
- MCP Server (The Provider): The execution plane that securely exposes the curated list of skills to your MCP-capable agents (n8n, Cursor, Claude Desktop, etc.).
๐ Requirements
- uv (Extremely fast Python package installer and resolver)
- Python 3.12+
- FastAPI
- FastMCP
- Agno
๐ ๏ธ Installation & Setup
-
Clone the repository:
git clone [https://github.com/chameleonbr/skills_mcp_server.git](https://github.com/chameleonbr/skills_mcp_server.git) cd skills_mcp_server -
Configure environment:
cp env.example .env # Edit .env and set API_KEY and SKILLS_DIR -
Install dependencies:
uv sync -
Run the server:
uv run uvicorn main:app --reload
Docker
docker-compose up --build
๐งช Running Tests
The project includes a comprehensive unit test suite covering models, routes, services, and the MCP server.
To run the tests, ensure you have the development dependencies installed, and then use pytest:
uv run pytest -v tests/
To run tests with coverage reporting:
uv run pytest --cov=. tests/
๐๏ธ Project Structure
skills_mcp_server/
โโโ main.py # FastAPI app + FastMCP mount + lifespan
โโโ routes.py # REST API routes (/skills)
โโโ services.py # SkillManager (install, delete, reload, GitHub support)
โโโ mcp_server.py # FastMCP server with Agno skill tools
โโโ models.py # Pydantic request/response schemas
โโโ s3_skills.py # S3Skills loader (sync from S3 bucket โ LocalSkills)
โโโ skills/ # Skill folders (each must contain SKILL.md)
โโโ pyproject.toml
โโโ Dockerfile / docker-compose.yml
๐ Authentication
All /skills routes require an X-API-Key header:
-H "X-API-Key: your_api_key"
Set API_KEY in your .env file.
๐ก REST API โ Control Plane
Interactive docs available at http://localhost:8000/docs.
Skills Management
| Method | Route | Description |
|---|---|---|
GET |
/skills |
List all loaded skills |
GET |
/skills/{name} |
Get full details of a skill |
POST |
/skills |
Install a skill from URL or base64 zip |
POST |
/skills/upload |
Install skill(s) via file upload |
DELETE |
/skills/{name} |
Delete a skill |
DELETE |
/skills |
Delete all skills |
GET |
/skills/prompt_snippet |
Get system prompt snippet for agents (supports ?skill_list=skill1,skill2 and ?prompt_enforcement=false) |
POST |
/skills/prompt_snippet |
Inject prompt snippet into the passed JSON payload (supports ?skill_list=... and ?prompt_enforcement=false) |
Health Check
GET /health
Returns 200 OK with the number of currently loaded skills.
๐ฅ Installing Skills
Note: You no longer need to provide a name. The server will automatically extract the skill name from the SKILL.md file inside the archive. If your folder or URL contains multiple skills, they will all be installed automatically.
1. File Upload (multipart)
curl -X POST http://localhost:8000/skills/upload \
-H "X-API-Key: your_api_key" \
-F "file=@my_skill.skill" \
-F "overwrite=true" # optional, defaults to false
2. URL (direct zip download)
curl -X POST "http://localhost:8000/skills" \
-H "X-API-Key: your_api_key" \
-H "Content-Type: application/json" \
-d '{"url": "https://example.com/my_skill.zip", "overwrite": true}'
3. Agent Skills Discovery RFC (Index JSON) ๐
Install multiple skills curated in a skills_index.json file as defined by the Cloudflare Discovery RFC:
curl -X POST "http://localhost:8000/skills" \
-H "X-API-Key: your_api_key" \
-H "Content-Type: application/json" \
-d '{"url": "https://raw.githubusercontent.com/LambdaTest/agent-skills/refs/heads/main/skills_index.json", "overwrite": true}'
3. GitHub URL ๐
You can point directly to a GitHub repository or a subdirectory inside one:
# Entire repository (uses main branch)
curl -X POST "http://localhost:8000/skills" \
-H "X-API-Key: your_api_key" \
-H "Content-Type: application/json" \
-d '{"url": "https://github.com/owner/repo"}'
# Specific branch
curl -X POST "http://localhost:8000/skills" \
-H "X-API-Key: your_api_key" \
-H "Content-Type: application/json" \
-d '{"url": "https://github.com/owner/repo/tree/develop"}'
# Specific subfolder inside a repo
curl -X POST "http://localhost:8000/skills" \
-H "X-API-Key: your_api_key" \
-H "Content-Type: application/json" \
-d '{"url": "https://github.com/anthropics/skills/tree/main/skills/skill-creator"}'
4. Base64 Zip
curl -X POST "http://localhost:8000/skills" \
-H "X-API-Key: your_api_key" \
-H "Content-Type: application/json" \
-d "{\"zip_base64\": \"$(base64 -w0 my_skill.zip)\", \"overwrite\": true}"
๐ค MCP Server โ Provider Plane
The MCP server is mounted at /mcp using the StreamableHTTP transport.
MCP Endpoint
| Transport | URL |
|---|---|
| StreamableHTTP | POST /mcp |
| SSE | GET /mcp/sse |
MCP Configuration (Claude Desktop / any MCP client)
{
"mcpServers": {
"skills": {
"url": "http://localhost:8000/mcp",
"headers": {
"X-API-Key": "your_api_key"
}
}
}
}
Available MCP Tools
| Tool | Description |
|---|---|
get_available_skills() |
List all loaded skills with name and description |
get_skill_instructions(skill_name) |
Load full instructions for a skill |
get_skill_reference(skill_name, reference_path) |
Read a reference document from a skill |
get_skill_script(skill_name, script_path, execute) |
Read or execute a script from a skill |
Agent System Prompt Integration
To inject skill awareness into any agent, retrieve the pre-built prompt snippet. You can optionally pass skill_list to only include specific skills by name. By default, a strict prompt enforcement string is prepended to ensure agents use tools exactly as instructed, which can be disabled via prompt_enforcement=false. The prompt enforcement string is regularly refined for optimal formatting and context efficiency.
# Get snippet containing ALL installed skills (includes prompt enforcement by default)
curl -X GET http://localhost:8000/skills/prompt_snippet \
-H "X-API-Key: your_api_key"
# Get snippet without the prompt enforcement rules
curl -X GET "http://localhost:8000/skills/prompt_snippet?prompt_enforcement=false" \
-H "X-API-Key: your_api_key"
# Get snippet filtered to specific skills
curl -X GET "http://localhost:8000/skills/prompt_snippet?skill_list=web_browsing,yahoo_finance" \
-H "X-API-Key: your_api_key"
Paste the returned XML snippet into your agent's system_prompt. It describes the loaded skills and how to use the MCP tools to access them.
Low-Code Integration (n8n, Make, Custom Apps)
To simplify integrations in platforms where manipulating strings is hard, use the POST endpoint sending the payload as JSON and receiving the same payload with the prompt with the additional system instructions that you need to append to the system prompt of your agent:
curl -X POST "http://localhost:8000/skills/prompt_snippet" \
-H "X-API-Key: your_api_key" \
-H "Content-Type: application/json" \
-d '{"messages": [{"role": "user", "content": "Help me"}], "temperature": 0.7}'
Returns the original payload with "prompt" dynamically appended/overwritten with the system instructions:
{
"messages": [{"role": "user", "content": "Help me"}],
"temperature": 0.7,
"prompt": "<skills_system>...</skills_system>"
}
๐ Skill Structure
Each skill is a folder inside SKILLS_DIR with the following layout:
my_skill/
โโโ SKILL.md # Required: frontmatter (name, description) + instructions
โโโ scripts/ # Optional: executable script files
โ โโโ run.py
โโโ references/ # Optional: documentation files
โโโ guide.md
SKILL.md example:
---
name: yahoo_finance
description: Fetch stock quotes and financial data from Yahoo Finance.
---
## Instructions
Use this skill when the user asks about stock prices or financial data...
โ๏ธ Environment Variables
Core
| Variable | Default | Description |
|---|---|---|
API_KEY |
(required) | Secret key for REST API and MCP authentication |
SKILLS_DIR |
skills |
Local directory for skill folders (used by local backend) |
SKILLS_STORAGE |
local |
Storage backend: local or s3 |
ALLOW_RUN_SCRIPTS |
false |
Whether to allow script execution via mcp_get_script |
LAZY_INSTALL_VENVS |
false |
If true, defers requirements.txt generation/installation to the first time a script is executed. If false, installs dependencies immediately upon skill installation. |
ALLOW_GET_AVAILABLE_SKILLS |
true |
Exposes the get_available_skills tool globally via FastMCP. Set to false to hide discovery mechanisms via tool calls. |
Detailed Explanation of Core Variables
API_KEY: This is the master secret used to secure both the REST API endpoints and the FastMCP transport connection. It is mandatory for any operation.SKILLS_DIR: Defines the path (relative to the app root or absolute) where the skill folders are maintained. Defaults toskillsin the local backend.SKILLS_STORAGE: Determines the strategy for loading skills. It can be set tolocal(reads fromSKILLS_DIR) ors3(syncs skills from an AWS S3-compatible object storage).ALLOW_RUN_SCRIPTS: A security toggle. If set tofalse(default), theget_skill_scripttool withexecute=Truewill be rejected by the application. Setting it totrueallows agents to effectively run python scripts defined by skills.LAZY_INSTALL_VENVS: Performance toggle. Iffalse(default), the server blocks during skill installation to build the Python Virtual Environment and install therequirements.txt. Iftrue, the installation is instantaneous, but the runtime will pause to install the.venvonly when a script from the skill is executed for the first time.ALLOW_GET_AVAILABLE_SKILLS: Integration toggle. Iftrue(default), theget_available_skillstool is exposed globally via FastMCP for dynamic discovery. Iffalse, the tool is hidden, requiring the system prompt to explicitly define the skills or using the/skills/prompt_snippetREST endpoint for injection, this is useful to avoid the LLM to discover skills that are not meant to be used, requiring the developer to explicitly define the skills in the system prompt, injecting the snippet.
S3 Storage (SKILLS_STORAGE=s3)
| Variable | Default | Description |
|---|---|---|
S3_BUCKET |
(required) | S3 bucket name |
S3_PREFIX |
skills/ |
Key prefix acting as the remote skills root |
S3_CACHE_DIR |
.s3cache |
Local directory where S3 files are cached |
AWS_ACCESS_KEY_ID |
โ | AWS access key (or use IAM role / ~/.aws/credentials) |
AWS_SECRET_ACCESS_KEY |
โ | AWS secret key |
AWS_DEFAULT_REGION |
us-east-1 |
AWS region |
AWS_ENDPOINT_URL |
โ | Custom endpoint for MinIO / LocalStack |
โ๏ธ Storage Backends
Local (default)
Skills are loaded from the SKILLS_DIR folder on disk.
With Docker, a named volume (skills_data) is used for persistence:
# .env
SKILLS_STORAGE=local
SKILLS_DIR=skills
The docker-compose.yml mounts skills_data:/app/skills automatically โ
skills survive docker-compose down / up cycles.
S3 / S3-compatible
Skills are synced from an S3 bucket to a local cache dir on every reload().
The S3Skills loader (s3_skills.py) follows the same SkillLoader interface as Agno's LocalSkills.
# .env
SKILLS_STORAGE=s3
S3_BUCKET=my-bucket
S3_PREFIX=skills/
AWS_ACCESS_KEY_ID=...
AWS_SECRET_ACCESS_KEY=...
AWS_DEFAULT_REGION=us-east-1
# MinIO / LocalStack:
# AWS_ENDPOINT_URL=http://localhost:9000
Expected bucket layout:
my-bucket/
โโโ skills/
โโโ my_skill/
โ โโโ SKILL.md
โ โโโ scripts/run.py
โ โโโ references/guide.md
โโโ another_skill/
โโโ SKILL.md
Recommended Servers
playwright-mcp
A Model Context Protocol server that enables LLMs to interact with web pages through structured accessibility snapshots without requiring vision models or screenshots.
Magic Component Platform (MCP)
An AI-powered tool that generates modern UI components from natural language descriptions, integrating with popular IDEs to streamline UI development workflow.
Audiense Insights MCP Server
Enables interaction with Audiense Insights accounts via the Model Context Protocol, facilitating the extraction and analysis of marketing insights and audience data including demographics, behavior, and influencer engagement.
VeyraX MCP
Single MCP tool to connect all your favorite tools: Gmail, Calendar and 40 more.
graphlit-mcp-server
The Model Context Protocol (MCP) Server enables integration between MCP clients and the Graphlit service. Ingest anything from Slack to Gmail to podcast feeds, in addition to web crawling, into a Graphlit project - and then retrieve relevant contents from the MCP client.
Kagi MCP Server
An MCP server that integrates Kagi search capabilities with Claude AI, enabling Claude to perform real-time web searches when answering questions that require up-to-date information.
E2B
Using MCP to run code via e2b.
Neon Database
MCP server for interacting with Neon Management API and databases
Qdrant Server
This repository is an example of how to create a MCP server for Qdrant, a vector search engine.
Exa Search
A Model Context Protocol (MCP) server lets AI assistants like Claude use the Exa AI Search API for web searches. This setup allows AI models to get real-time web information in a safe and controlled way.