Mythosia.AI.Mcp 0.0.1-preview

Mythosia.AI.Mcp

MCP (Model Context Protocol) client integration for Mythosia.AI.
Connect to any MCP server and automatically register its tools as FunctionDefinitions — usable by all AI providers (OpenAI, Anthropic, Google, xAI, DeepSeek, etc.).

Installation

dotnet add package Mythosia.AI.Mcp

Quick Start

using Mythosia.AI.Mcp;
using Mythosia.AI.Services.OpenAI;

var service = new OpenAIService(apiKey, httpClient);

// Connect to an MCP server (stdio transport)
await using var mcp = await service.WithMcpServerAsync(
    "npx", "-y @modelcontextprotocol/server-filesystem /workspace");

// AI now has access to all MCP tools automatically
var answer = await service.GetCompletionAsync("List all files in /workspace");

Features

• Automatic tool discovery — tools/list is called on connect; each tool becomes a FunctionDefinition
• Provider-agnostic — MCP tools work with OpenAI, Anthropic, Google, xAI, DeepSeek, Perplexity
• Composable — mix MCP tools with [AiFunction] local functions and FunctionBuilder definitions
• Lifecycle management — IAsyncDisposable pattern; await using shuts down the server cleanly
• Tool filtering — include only the tools you need with a name filter
• Name prefixing — avoid tool name collisions when connecting multiple MCP servers

Usage

Stdio Transport (most common)

// Basic — command + args
await using var mcp = await service.WithMcpServerAsync(
    "npx", "-y @modelcontextprotocol/server-filesystem /workspace");

// With environment variables
await using var mcp = await service.WithMcpServerAsync(
    "python", "-m my_mcp_server",
    environmentVariables: new Dictionary<string, string>
    {
        ["API_KEY"] = "sk-..."
    });

Tool Filtering

// Only include specific tools
await using var mcp = await service.WithMcpServerAsync(
    "npx", "-y @modelcontextprotocol/server-github",
    toolFilter: name => name is "search_repositories" or "get_file_contents");

Multiple MCP Servers

// Connect to multiple servers — use namePrefix to avoid collisions
await using var fs = await service.WithMcpServerAsync(
    "npx", "-y @modelcontextprotocol/server-filesystem /workspace",
    namePrefix: "fs_");

await using var gh = await service.WithMcpServerAsync(
    "npx", "-y @modelcontextprotocol/server-github",
    namePrefix: "gh_");

// AI sees tools like: fs_read_file, fs_write_file, gh_search_repositories, etc.

Custom Transport

using Mythosia.AI.Mcp.Transports;

// Use any IMcpTransport implementation
var transport = new StdioTransport("node", "my-server.js");
await using var mcp = await service.WithMcpServerAsync(transport);

Direct McpConnection Usage

// For lower-level control without AIService integration
var transport = new StdioTransport("npx", "-y @modelcontextprotocol/server-filesystem /tmp");
var connection = new McpConnection(transport);
await connection.InitializeAsync();

// Inspect discovered tools
foreach (var tool in connection.Tools)
    Console.WriteLine($"{tool.Name}: {tool.Description}");

// Call a tool directly
var result = await connection.CallToolAsync("read_file", 
    new Dictionary<string, object> { ["path"] = "/tmp/hello.txt" });

// Convert to FunctionDefinitions manually
var functions = McpToolAdapter.ToFunctionDefinitions(connection);

Architecture

Mythosia.AI.Abstractions   ← FunctionDefinition, FunctionParameters
        ↑
    Mythosia.AI             ← AIService, provider implementations
        ↑
  Mythosia.AI.Mcp           ← McpConnection, StdioTransport, McpToolAdapter

MCP tools are converted to FunctionDefinition at connection time. The existing function calling infrastructure (ReAct loop, tool execution) handles the rest — no changes needed in the core package.

Supported MCP Protocol Features

Links

• Mythosia.AI Documentation
• MCP Specification
• GitHub