SkillMCP (Skill Management System)

A containerized, horizontally scalable Model Context Protocol (MCP) server for distributing and managing AI agent skills.

Overview & Motivation

AI agents rely on domain-specific skills (instructions, metadata, schemas, and reference assets) to perform complex engineering and analytical tasks. However, managing skills across diverse teams and agent fleets often introduces critical operational challenges:

• Fragmented & Outdated Skills: Skills stored across scattered individual repositories or copied manually quickly fall out of sync, leaving agents executing obsolete or incompatible workflows.

• Distribution & Update Bottlenecks: Distributing skill updates across distributed agent instances requires manual synchronization or fragile file-copy steps.

• Lack of Versioning & Diagnostics Friction: When skills are edited without immutable versioning, diagnosing regressions or agent behavior shifts becomes nearly impossible.

• Single-Host Scalability Limits: Traditional stdio-based MCP servers are tied to single local host processes, blocking horizontal scaling and high availability.

SkillMCP solves these challenges by providing a centralized, containerized, and horizontally scalable Skill Management System powered by Stateless Streamable HTTP.

Related MCP server: SkillShare MCP Server

Key Capabilities & Features

1. Centralized & Versioned Skill Packaging

• Immutable Container Releases: Skills (SKILL.md, references/, and examples/) are packaged directly inside Docker images tagged with explicit semver (v1.2.0), ensuring 100% reproducible environments and auditability.

• Unified Skill Repository: Eliminates fragmented multi-repo drift by managing, validating, and bundling all domain skills in a single maintainable repository.

• Fast Troubleshooting & Traceability: Versioned container tags make it straightforward to diagnose agent issues, reproduce historical behavior, and roll back changes instantly.

2. Stateless MCP over Streamable HTTP (Horizontal Scalability & HPA Ready)

• True Stateless Request/Response Architecture: Streamable HTTP uses standard HTTP POST requests where backend instances do not maintain long-lived in-memory socket state between client calls.

• Seamless Horizontal Pod Autoscaling (HPA): Without sticky sessions or persistent TCP stream locking, backend replicas can scale up/down dynamically and handle requests evenly across any load balancer.

• Short-Lived Streaming: Responses requiring streaming are upgraded to text/event-stream only for the duration of that specific payload and close immediately once the JSON-RPC response finishes.

• Nginx Ingress Load Balancing: Configured with least_conn routing, keepalive connection pooling, and dedicated /healthz health checks for zero-downtime rolling updates.

3. Developer & Agent Tooling

• Built-in Skill Validator CLI: skillmcp validate ./skills automatically verifies directory structures, YAML frontmatter, and asset links before packaging.

• Dynamic Discovery & Search:

  • MCP Tools: list_skills, get_skill, search_skills, read_skill_reference, read_skill_example.

  • MCP Resources: skill://{name} for direct markdown document inspection.

• Dual Compose Environments: docker-compose.local.yml for instant local development with volume-mounted hot reloading, and docker-compose.yml for production deployments.

• TREM Python Standard: Built strictly following Testable, Readable, Extensible, and Maintainable (TREM) principles with uv, pydantic-settings, standard library logging, and pytest.

• Automated CI/CD Publishing: GitHub Actions pipeline that validates tests and pushes immutable semver releases to Docker Hub on version tags (v*.*.*).

Quickstart

Prerequisites

• Python 3.11+

• uv

• Docker & Docker Compose

Local Installation

# Sync dependencies
uv sync

# Run tests
uv run pytest -v

# Validate skills
uv run skillmcp validate ./skills

# List discovered skills
uv run skillmcp list --skills-path ./skills

Running the Server Locally

# Start MCP server directly (Streamable HTTP on port 8000)
uv run skillmcp serve --host 0.0.0.0 --port 8000

Container Usage & Deployment

1. Run Standalone Docker Container

Pull and run the pre-built image directly from Docker Hub:

# Run standalone container with bundled skills
docker run -d \
  --name skillmcp \
  -p 8000:8000 \
  docker.io/clivechung/skillmcp:latest

# Or mount your own custom skills directory
docker run -d \
  --name skillmcp \
  -p 8000:8000 \
  -v $(pwd)/skills:/app/skills:ro \
  docker.io/clivechung/skillmcp:latest

Verify the server is running:

curl http://localhost:8000/healthz
# {"status":"healthy","service":"skillmcp","version":"0.1.0"}

2. Run with Docker Compose (Production Topology)

Runs 2 backend skillmcp replicas behind an Nginx load balancer:

docker compose up -d

• MCP Endpoint: http://localhost:8080/mcp

• Ingress Health Check: http://localhost:8080/healthz

3. Local Development (Live Reload & Volume Mounts)

docker compose -f docker-compose.local.yml up -d --build

• Nginx Ingress: http://localhost:8080/mcp

• Backend App (Direct): http://localhost:8000/mcp

Client Integration Guide

SkillMCP exposes a stateless Model Context Protocol (MCP) server over Streamable HTTP at /mcp. Configure your favorite AI coding assistant or agent CLI using the examples below.

1. Google Antigravity (AGY)

Add SkillMCP to your Antigravity configuration (either workspace-level .agents/mcp_config.json or global ~/.gemini/config/mcp_config.json):

{
  "mcpServers": {
    "skillmcp": {
      "url": "http://localhost:8080/mcp"
    }
  }
}

Note: If connecting directly to the standalone container without Nginx, use http://localhost:8000/mcp.

2. OpenAI Codex (Visio IDE)

For projects developed in Visio / VS Code IDE with OpenAI Codex, add the server to your project's .vscode/mcp.json or workspace settings:

{
  "mcpServers": {
    "skillmcp": {
      "url": "http://localhost:8080/mcp"
    }
  }
}

If your IDE environment or extension utilizes a stdio bridge for remote HTTP endpoints, configure mcp-remote:

{
  "mcpServers": {
    "skillmcp": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:8080/mcp"]
    }
  }
}

3. Claude CLI (Visio IDE & Terminal)

Direct Registration via Claude CLI:

In your Visio IDE integrated terminal or command line:

# Add the streamable HTTP MCP server to Claude CLI
claude mcp add --transport http skillmcp http://localhost:8080/mcp

Via Visio / Claude MCP Configuration (~/.claude.json or claude_desktop_config.json):

{
  "mcpServers": {
    "skillmcp": {
      "url": "http://localhost:8080/mcp"
    }
  }
}

Or using mcp-remote for stdio-only bridge clients:

{
  "mcpServers": {
    "skillmcp": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "http://localhost:8080/mcp"]
    }
  }
}

Available MCP Tools & Resources

Once connected, your agents will have immediate access to the following tools:

Testing Seams

• Seam 1: Domain Service: tests/test_domain_service.py (Validates scanner, parser, traversal safety, and query engine)

• Seam 2: MCP Protocol & Tools: tests/test_mcp_server.py & tests/test_mcp_http.py (Validates FastMCP tools, resources, and ASGI transport routes)

• Seam 3: CLI & Integration: tests/test_cli.py (Validates CLI validator, list, and serve commands)

License & Attribution

• Core Project: Released under the MIT License (c) 2026 clivechung.

• Skills & Attributions: Declarations, licenses, and provenance for all bundled server skills and agent development skills are documented in THIRD_PARTY_NOTICES.md.