maas-mcp-server
Enables AI assistants to interact with MAAS (Metal as a Service) infrastructure for machine management, network configuration, and storage management through a standardized JSON-RPC 2.0 interface.
README
MAAS MCP Server
A Model Context Protocol (MCP) server for interacting with MAAS (Metal as a Service) through a standardized JSON-RPC 2.0 interface.
Introduction
The MAAS MCP Server implements the Model Context Protocol specification to enable AI assistants to interact with MAAS (Metal as a Service) infrastructure. It provides a standardized interface for machine management, network configuration, storage management, and more through a JSON-RPC 2.0 API.
This server can operate in two transport modes:
- HTTP/HTTPS - For web-based interactions and SSE (Server-Sent Events) streaming
- stdin/stdout - For direct integration with AI assistants and CLI tools
Installation
Prerequisites
- Go 1.22 or later
- MAAS server with API access
- MAAS API key in the format "consumer:token:secret"
With script from release
curl -sSL https://raw.githubusercontent.com/lspecian/maas-mcp-server/main/scripts/install.sh | bash
From Source
-
Clone the repository:
git clone https://github.com/lspecian/maas-mcp-server.git cd maas-mcp-server -
Build the server:
./build.sh build -
Configure the server:
cp config/config.yaml.example config/config.yaml cp .env.example .env -
Edit
config/config.yamland update:- MAAS API URL
- MAAS API key
- Server host/port
- Logging configuration
Environment Variables
The server requires the following environment variables:
MAAS_API_URL- The URL of your MAAS API endpointMAAS_API_KEY- Your MAAS API key in the format "consumer:token:secret"LOG_LEVEL- (Optional) The logging level (default: "info")LOG_FORMAT- (Optional) The logging format (default: "json")AUTH_ENABLED- (Optional) Whether authentication is enabled (default: "false")
These can be set in the .env file or directly in your environment.
Roo Integration with .roo/mcp.json
For integration with Roo, you can configure the server using the .roo/mcp.json file. Here's an example configuration:
{
"mcpServers": {
"maas-server": {
"command": "./maas-mcp-server",
"args": [
"stdio"
],
"protocol": "stdio",
"jsonrpc": "2.0",
"readyMessage": "MCP server ready",
"env": {
"MAAS_API_URL": "http://your-maas-server:5240/MAAS",
"MAAS_API_KEY": "consumer:token:secret",
"LOG_LEVEL": "debug",
"LOG_FORMAT": "json",
"AUTH_ENABLED": "false"
},
"disabled": false,
"alwaysAllow": [
"maas_list_machines",
"maas_get_machine_details",
"maas_power_on_machine",
"maas_power_off_machine",
"list_machines"
]
}
}
}
The environment variables in the env section are used to configure the server. The server will read these variables and use them to configure itself.
Using Docker
-
Build the Docker image:
./build-docker.sh -
Run the container:
docker run -p 8081:8081 -v $(pwd)/config:/app/config maas-mcp-server
Usage
Starting the Server
HTTP Mode
./build.sh run
The server will be available at http://localhost:8081/mcp.
stdin/stdout Mode
./build.sh run-mcp-stdio
Or directly:
./maas-mcp-server stdio
In this mode, the server reads JSON-RPC requests from stdin and writes responses to stdout, making it suitable for integration with AI assistants.
Version Information
To display the version information:
./maas-mcp-server --version
JSON-RPC 2.0 Message Format
The server communicates using the JSON-RPC 2.0 protocol. Here's the basic message format:
Request
{
"jsonrpc": "2.0",
"method": "method_name",
"params": {
"param1": "value1",
"param2": "value2"
},
"id": "request-id"
}
Response (Success)
{
"jsonrpc": "2.0",
"result": {
"key1": "value1",
"key2": "value2"
},
"id": "request-id"
}
Response (Error)
{
"jsonrpc": "2.0",
"error": {
"code": -32000,
"message": "Error message",
"data": {}
},
"id": "request-id"
}
Discovery
To discover the capabilities of the MCP server, send a discovery request:
{
"jsonrpc": "2.0",
"method": "discover",
"params": {},
"id": "1"
}
The server will respond with information about available tools and resources:
{
"jsonrpc": "2.0",
"result": [
{
"name": "maas_list_machines",
"description": "List all machines managed by MAAS with filtering and pagination",
"parameters": {
"type": "object",
"properties": {
"hostname": { "type": "string" },
"zone": { "type": "string" },
"pool": { "type": "string" },
"status": { "type": "string" },
"power_state": { "type": "string" },
"system_id": { "type": "string" },
"architecture": { "type": "string" },
"tags": { "type": "array", "items": { "type": "string" } },
"limit": { "type": "integer" },
"page": { "type": "integer" } // Example parameter, actual parameters are dynamically generated
}
}
}
// ... (list of all dynamically generated tools)
],
"id": "1"
}
The list of tools is dynamically generated based on the MAAS API documentation. This provides a comprehensive set of tools for interacting with various MAAS functionalities. For details on how these tools are generated and how to update them, please refer to the MAAS API Tool Generation documentation.
Example Workflows
Listing Available Machines
package main
import (
"bufio"
"encoding/json"
"fmt"
"io"
"os"
"os/exec"
)
func main() {
// Start the MCP server in stdio mode
cmd := exec.Command("./build.sh", "run-mcp-stdio")
stdin, _ := cmd.StdinPipe()
stdout, _ := cmd.StdoutPipe()
cmd.Start()
// Wait for server to start
reader := bufio.NewReader(stdout)
for {
line, _ := reader.ReadString('\n')
if line == "MCP server ready\n" {
break
}
}
// Send list machines request
request := map[string]interface{}{
"jsonrpc": "2.0",
"method": "maas_list_machines",
"params": map[string]interface{}{},
"id": "1",
}
requestJSON, _ := json.Marshal(request)
io.WriteString(stdin, string(requestJSON)+"\n")
// Read response
responseStr, _ := reader.ReadString('\n')
var response map[string]interface{}
json.Unmarshal([]byte(responseStr), &response)
// Pretty print the response
prettyJSON, _ := json.MarshalIndent(response, "", " ")
fmt.Println(string(prettyJSON))
cmd.Process.Kill()
}
Getting Machine Details and Powering On
package main
import (
"bufio"
"encoding/json"
"fmt"
"io"
"os"
"os/exec"
)
func main() {
// Start the MCP server in stdio mode
cmd := exec.Command("./build.sh", "run-mcp-stdio")
stdin, _ := cmd.StdinPipe()
stdout, _ := cmd.StdoutPipe()
cmd.Start()
// Wait for server to start
reader := bufio.NewReader(stdout)
for {
line, _ := reader.ReadString('\n')
if line == "MCP server ready\n" {
break
}
}
// Get machine details
detailsRequest := map[string]interface{}{
"jsonrpc": "2.0",
"method": "maas_get_machine_details",
"params": map[string]string{"system_id": "abc123"},
"id": "1",
}
requestJSON, _ := json.Marshal(detailsRequest)
io.WriteString(stdin, string(requestJSON)+"\n")
// Read response
responseStr, _ := reader.ReadString('\n')
var response map[string]interface{}
json.Unmarshal([]byte(responseStr), &response)
// Check if machine is powered off
result := response["result"].(map[string]interface{})
if result["power_state"] == "off" {
// Power on the machine
powerRequest := map[string]interface{}{
"jsonrpc": "2.0",
"method": "maas_power_on_machine",
"params": map[string]string{"system_id": "abc123"},
"id": "2",
}
requestJSON, _ = json.Marshal(powerRequest)
io.WriteString(stdin, string(requestJSON)+"\n")
// Read power on response
powerResponseStr, _ := reader.ReadString('\n')
fmt.Println(powerResponseStr)
}
cmd.Process.Kill()
}
Building and Testing
Building from Source
# Build the server
./build.sh build
# Build the clean architecture version
./build.sh build-mcp
Running Tests
# Run all tests
./build.sh test
# Run specific tests
go test -v ./internal/service/...
Integration Tests
# Run the test client
go run test/go/test-stdio-client.go
Releases
The project uses GitHub Actions to automatically build and publish binaries for multiple platforms when a new tag is pushed to the repository.
Automatic Release Process
- Update the version in
internal/version/version.go - Update the
CHANGELOG.mdwith the changes in the new version - Commit the changes
- Create and push a new tag:
git tag v1.0.0 git push origin v1.0.0 - The GitHub Actions workflow will automatically:
- Build binaries for Linux (amd64, arm64), macOS (amd64, arm64), and Windows (amd64)
- Create a GitHub Release with the tag name
- Upload the binaries as assets
- Generate SHA256 checksums for all binaries
- Add release notes based on the CHANGELOG.md
Manual Release Process
You can also trigger a manual release using the GitHub Actions workflow:
- Go to the GitHub repository
- Click on the "Actions" tab
- Select the "Manual Release" workflow
- Click on "Run workflow"
- Enter the version number (e.g., "1.1.0")
- Select whether this is a pre-release
- Click "Run workflow"
The workflow will:
- Build binaries for all supported platforms
- Create a GitHub Release with the specified version
- Upload the binaries as assets
- Generate SHA256 checksums for all binaries
- Add release notes from the CHANGELOG.md
Binary Naming Convention
Binaries follow a consistent naming pattern:
maas-mcp-server-{version}-{os}-{arch}[.exe]
For example:
maas-mcp-server-1.0.0-linux-amd64maas-mcp-server-1.0.0-linux-arm64maas-mcp-server-1.0.0-darwin-amd64maas-mcp-server-1.0.0-darwin-arm64maas-mcp-server-1.0.0-windows-amd64.exe
Building for Different Platforms
You can build the server for different platforms using the Go cross-compilation feature:
# Build for Linux AMD64
GOOS=linux GOARCH=amd64 go build -o maas-mcp-server-linux-amd64 pkg/mcp/cmd/main.go
# Build for Linux ARM64
GOOS=linux GOARCH=arm64 go build -o maas-mcp-server-linux-arm64 pkg/mcp/cmd/main.go
# Build for macOS AMD64
GOOS=darwin GOARCH=amd64 go build -o maas-mcp-server-darwin-amd64 pkg/mcp/cmd/main.go
# Build for macOS ARM64 (Apple Silicon)
GOOS=darwin GOARCH=arm64 go build -o maas-mcp-server-darwin-arm64 pkg/mcp/cmd/main.go
# Build for Windows AMD64
GOOS=windows GOARCH=amd64 go build -o maas-mcp-server-windows-amd64.exe pkg/mcp/cmd/main.go
API Reference
Error Codes
| Code | Message | Description |
|---|---|---|
| -32700 | Parse error | Invalid JSON was received |
| -32600 | Invalid request | The JSON sent is not a valid Request object |
| -32601 | Method not found | The method does not exist / is not available |
| -32602 | Invalid params | Invalid method parameter(s) |
| -32603 | Internal error | Internal JSON-RPC error |
| -32000 | Authentication failed | Failed to authenticate with MAAS |
| -32001 | Rate limit exceeded | Too many requests |
| -32002 | Version not supported | The requested MCP version is not supported |
| -32003 | Resource not found | The requested resource was not found |
| -32004 | Operation failed | The requested operation failed |
Available Methods
The available methods (tools) are discovered via the discover JSON-RPC method. The server dynamically registers a comprehensive set of tools based on the parsed MAAS API. Refer to the "Discovery" section above for an example of how to retrieve the list of available tools and their schemas.
Tool Generation
The MAAS tools provided by this server are dynamically generated from the MAAS API documentation. This ensures that the server can adapt to a wide range of MAAS API functionalities. For detailed information on how these tools are parsed, generated, and how to update them if the MAAS API changes, please see the MAAS API Tool Generation documentation in cmd/gen-tools/README.md.
Contributing
Contributions are welcome! Please see CONTRIBUTING.md for details.
License
This project is licensed under the MIT License - see the LICENSE file for details.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。