MCP Platform

A self-hosted platform for creating, deploying and managing MCP (Model Context Protocol) servers — without writing code.

Connect AI assistants (Continue, OpenWebUI, Claude Desktop) to your infrastructure: databases, APIs, Kubernetes clusters, Linux servers — all through a visual web UI.

---

What it does

• Visual tool builder — create MCP tools from a web form, not code
• Instant deploy — one click deploys a containerized MCP server
• Multiple runtime types — shell commands, HTTP APIs, custom images
• Policy engine — allowlist binaries, block dangerous commands, enforce read-only mode
• OpenWebUI integration — Tool Server (OpenAPI), Python tool import, filter pipeline
• Continue integration — streamable-http transport, auto-generated config snippets
• Audit log — every tool invocation logged with caller IP and model name

---

Prerequisites

Note: Docker Desktop on macOS/Windows works for development but runtime containers may not be reachable by AI clients outside the VM. Linux is recommended for production.

Install Docker (Ubuntu)

curl -fsSL https://get.docker.com | sh
sudo usermod -aG docker $USER
newgrp docker
docker --version        # Docker version 24.x.x
docker compose version  # Docker Compose version v2.x.x

Important: Your user must be in the docker group, otherwise install.sh and all docker commands will fail with permission errors.

# Add current user to docker group
sudo usermod -aG docker $USER

# Apply without logout (current session only)
newgrp docker

# Verify
docker ps   # should work without sudo

If you log out and back in, group membership is applied automatically.

---

Installation

# 1. Install Docker (if not already installed)
curl -fsSL https://get.docker.com | sh
sudo usermod -aG docker $USER
newgrp docker

# 2. Clone the repository
git clone https://github.com/YOUR_USERNAME/mcp-platform.git
cd mcp-platform

# 3. Run the installer
chmod +x install.sh
./install.sh

Open http://YOUR_SERVER_IP:18100 — login: admin / admin

Updating to a newer version

git pull
docker compose up -d --build mcp-platform mcp-platform-operator

Note: --force-recreate alone does not rebuild images from updated source code.
Always use --build after git pull.

Rebuilding images manually (after local code changes)

# Rebuild control plane only
docker compose up -d --build mcp-platform

# Rebuild operator only
docker compose up -d --build mcp-platform-operator

# Rebuild runtime images (shell / http-gateway)
docker compose --profile build-only build
# Then redeploy running runtimes via the UI: Runtime → Deploy

# Rebuild everything at once
docker compose build && docker compose up -d

---

Quick Start

Requirements: Docker 24+, Docker Compose v2, Linux

git clone https://github.com/YOUR_USERNAME/mcp-platform.git
cd mcp-platform
chmod +x install.sh
./install.sh

Open http://localhost:18100 — login: admin / admin

---

Manual Setup

# 1. Copy and edit configuration
cp .env.example .env
# Edit MCP_HOST_DATA_PATH to the absolute path of the data/ folder

# 2. Build runtime images
docker compose --profile build-only build

# 3. Start the platform
docker compose up -d

# 4. Open the UI
open http://localhost:18100

---

Architecture

┌─────────────────────────────────────────────────┐
│                   MCP Platform UI                │
│            (FastAPI · SQLite · port 18100)        │
└────────────────────┬────────────────────────────┘
                     │ manages
         ┌───────────▼──────────┐
         │   Operator (worker)   │
         │   Docker SDK          │
         └───────────┬──────────┘
                     │ creates/stops
     ┌───────────────┼───────────────────┐
     ▼               ▼                   ▼
┌─────────┐   ┌─────────────┐   ┌──────────────┐
│ Shell   │   │ HTTP Gateway│   │ Custom Image │
│ Runtime │   │ Runtime     │   │ Runtime      │
│ /mcp    │   │ /mcp        │   │ /mcp         │
└─────────┘   └─────────────┘   └──────────────┘
     ▲               ▲
     │ calls         │ calls
┌────┴───────────────┴──────┐
│  AI Client                │
│  Continue / OpenWebUI     │
└───────────────────────────┘

---

Base Runtime Images

Every MCP runtime container must use one of the two platform base images (or a custom image built on top of them). These images contain the built-in MCP server (FastAPI + uvicorn) that handles /mcp, /tools/{name}, /openwebui and /health endpoints.

mcp-runtime-shell:latest

Built from runtime-shell/ during docker compose build.

Base: python:3.12-slim (Debian Bookworm Slim)

Included tools:

Use for: shell tools — curl, oc, kubectl, psql, ping, ssh, any CLI command

mcp-runtime-http-gateway:latest

Built from runtime-http-gateway/ during docker compose build.

Base: python:3.12-slim (Debian Bookworm Slim)

Included:

Use for: HTTP tools — calling external REST APIs, webhooks, integrations

Building a custom image

If you need tools not available in the base images (e.g. psql, terraform, awscli), use the Image Builder in the UI or write your own Dockerfile:

# ALWAYS start from a platform base image — never from a raw OS image
FROM mcp-runtime-shell:latest

RUN apt-get update && apt-get install -y postgresql-client && rm -rf /var/lib/apt/lists/*

⚠️ Raw OS images will NOT work.
ubuntu:24.04, debian:bookworm-slim, alpine etc. do not contain the MCP server.
The container starts with bash, exits immediately, and enters a restart loop.

---

Creating your first MCP server

Option A — Use an example package

1. Go to Tool Packages → Import Package JSON
2. Upload a file from examples/ (e.g. curl-http-toolkit.json)
3. Click Create MCP → name it → Deploy
4. Copy the endpoint URL → paste into Continue or OpenWebUI

Option B — Build from scratch

1. Tool Packages → + Generate Package
2. Fill in name, choose runtime type (Shell or HTTP)
3. Add tools with commands/URLs and input parameters
4. Install → Create MCP Server → Deploy

See docs/jak-stworzyc-mcp-server.md for full step-by-step guide.

Option C — Let AI create it for you (Platform Manager)

MCP Platform includes a built-in Platform Manager MCP server — a meta-MCP that lets AI models create other MCP servers automatically.

1. Go to Quick Start → deploy MCP Platform Manager
2. Add credential: PLATFORM_URL = http://mcp-platform:8080
3. Deploy → connect to your AI client
4. Tell the AI what you need:

"Create an MCP server for monitoring MikroTik router at 192.168.88.1 with login admin and password test123"

The AI will:

• Read platform instructions (get_instructions)
• Generate the package JSON with tools, policy, and credentials
• Call create_mcp_server to deploy it automatically

Available tools in Platform Manager:

Built-in template packages

---

Connect to Continue

{
  "mcpServers": {
    "my-server": {
      "url": "http://localhost:PORT/mcp",
      "transport": "streamable-http"
    }
  }
}

OpenWebUI Integration

MCP Platform has native integration with OpenWebUI.
Each deployed MCP server automatically exposes three integration endpoints:

Option 1 — Tool Server (recommended, multiple tools at once)

Each runtime serves a full OpenAPI spec at /openwebui/openapi.json.
OpenWebUI discovers all tools automatically.

1. OpenWebUI → Admin Panel → Tools → Tool Servers → Add
2. Enter URL: http://YOUR_IP:PORT/openwebui
3. Save — all tools from the MCP server appear instantly

OpenWebUI automatically appends /openapi.json — enter the base URL only.

Option 2 — Python Tool (import from link)

Each runtime generates a ready-to-import Python tool file.

1. Open the runtime detail page → Podłącz tab → copy the Python Tool link
2. OpenWebUI → Workspace → Tools → Import from URL → paste the link
3. The tool appears in your workspace with full function signatures

Option 3 — RAGHybrid filter pipeline

If you run RAGHybrid alongside MCP Platform, use the auto-context filter:

• Automatically injects RAG context into prompts when the tool is active
• Configurable top_k, retrieval mode, and source filtering

---

Example Packages

---

Configuration

---

License

Copyright © 2026 — All Rights Reserved.
See LICENSE MIT