Connect Delphi applications to the AI ecosystem through the Model Context Protocol: TsgcWSAPIServer_MCP exposes your application's tools, prompts and resources to AI models, and TsgcWSAPIClient_MCP consumes any MCP-compatible server. The same library ships AI API clients for OpenAI, Anthropic Claude and Google Gemini.
MCP is the emerging standard for connecting large language models to external tools and data. Instead of one custom integration per AI provider, you implement one protocol.
Model Context Protocol in Delphi means your existing VCL or FMX application can become an AI-accessible service. The protocol is JSON-RPC based and defines three primitives: tools (callable functions with typed parameters), prompts (reusable templates with arguments) and resources (data exposed through URI addressing). An MCP server publishes them; an MCP client (Claude Desktop, Cursor, or your own Delphi application) discovers and calls them.
The server component attaches to an sgcWebSockets HTTP server and supports both stdio and HTTP transports, so the same code serves a desktop AI assistant launching it as a subprocess and a remote agent calling it over the network. The client component talks JSON-RPC over HTTP or HTTP Streamable and covers the full catalogue surface: initialize, list and call tools, retrieve prompts, read resources, plus sampling and elicitation for interactive AI workflows.
Both components implement the MCP 2026-07-28 specification: stateless requests with per-request _meta, server/discover, subscriptions/listen, Multi Round-Trip Requests and the Tasks extension. They stay backward compatible with 2025-11-25. The same server answers both eras, and the client falls back automatically when a server only speaks the older protocol.
TsgcWSAPIServer_MCP: tools, prompts, resources, stdio and HTTP transports
TsgcWSAPIClient_MCP: discover and call any MCP-compatible server
TsgcHTTP_API_OpenAI, TsgcHTTP_API_Anthropic, TsgcHTTP_API_Gemini
Delphi 7 through RAD Studio 13, C++Builder 10.1 Berlin through 13, Lazarus 4.4.0
Server, client and the AI provider clients that consume what MCP exposes.
Tools.AddTool registers a callable function and its InputSchema properties; calls arrive in the OnMCPRequestTool event with parsed arguments and a structured response object.
Reusable prompt templates with arguments and URI-addressed resources, served through OnMCPRequestPrompt and OnMCPRequestResource.
Initialize, ListTools, ListPrompts, ListResources, ListResourceTemplates, then RequestTool / RequestPrompt / RequestResource with event-driven responses.
The client supports MCP sampling (server-requested AI model interaction) and elicitation (gathering user input mid-workflow), the two interactive extensions of the protocol.
Client-side authentication options include API keys via AuthenticationOptions.ApiKey; see the blog for OAuth2-protected MCP servers.
Stateless requests, server/discover, Mcp-Method / Mcp-Name / Mcp-Param headers with validation, cache hints (ttlMs, cacheScope) and the new error codes. MCPOptions.ProtocolEra on the client selects Auto, Modern or Legacy.
A tool asks for elicitation (form or URL), sampling or roots through aResponse.InputRequired; the client answers in OnMCPInputRequired and retries with the HMAC-signed requestState.
subscriptions/listen streams list changes and resource updates. The io.modelcontextprotocol/tasks extension (CreateTask) runs long tools in the background, and the client component polls it with MCPOptions.Tasks.Enabled or fetches it directly with TasksGet.
MCPOptions.Authorization: protected resource metadata discovery, Client ID Metadata Documents, dynamic registration fallback, PKCE with resource indicators, RFC 9207 iss validation and per-issuer credentials.
Stdio for subprocess-launched servers (the Claude Desktop model) and HTTP for network deployments. One component, both channels.
Talk to the models directly too: chat, streaming and function calling against OpenAI, Anthropic Claude and Google Gemini, plus DeepSeek, Grok, Mistral and Ollama in the same AI family.
Windows 32/64, Linux 64, macOS (Intel and ARM), iOS and Android. VCL and FireMonkey, with design-time components.
The MCP server, MCP client and AI API clients are Enterprise edition features of sgcWebSockets.
Register a tool with a typed argument on the server; discover and call it from the client.
uses
sgcWebSocket_Server, sgcAI, sgcAI_MCP_Classes, sgcAI_MCP_Server;
procedure TForm1.SetupMCPServer;
begin
// Attach the MCP API component to a sgcWebSockets HTTP server
MCPServer.Server := Server;
MCPServer.EndpointOptions.Endpoint := '/mcp';
MCPServer.MCPOptions.ServerInfo.Name := 'sgc-mcp-server';
MCPServer.MCPOptions.ServerInfo.Version := '1.0.0';
// Register a callable tool with a typed argument
with MCPServer.Tools.AddTool('GetTemperature',
'Get the actual temperature in a city.') do
InputSchema.Properties.AddProperty('city', True);
MCPServer.OnMCPRequestTool := MCPRequestTool;
Server.Port := 8080;
Server.Active := True;
end;
procedure TForm1.MCPRequestTool(Sender: TObject;
const aSession: TsgcAI_MCP_Session;
const aRequest: TsgcAI_MCP_Request_ToolsCall;
const aResponse: TsgcAI_MCP_Response_ToolsCall);
begin
if aRequest.Params.Name = 'GetTemperature' then
aResponse.Result.Content.AddText('The current temperature in ' +
aRequest.Params.Arguments.Item[0].Value + ' is 22 Celsius');
end;uses
sgcAI_MCP_Client, sgcAI_MCP_Classes;
var
MCP: TsgcWSAPIClient_MCP;
begin
MCP := TsgcWSAPIClient_MCP.Create(nil);
MCP.MCPOptions.HttpOptions.URL := 'https://mcp.example.com/';
MCP.MCPOptions.ClientInfo.Name := 'sgc-mcp-client';
MCP.MCPOptions.ClientInfo.Title := 'sgc MCP demo';
MCP.MCPOptions.ClientInfo.Version := '1.0.0';
MCP.MCPOptions.AuthenticationOptions.ApiKey.Enabled := True;
MCP.MCPOptions.AuthenticationOptions.ApiKey.Value := 'sk-mcp-...';
MCP.OnMCPInitialize := MCPInit;
MCP.OnMCPListPrompts := MCPListPrompts;
MCP.OnMCPListTools := MCPListTools;
MCP.OnMCPResponseTool := MCPToolResponse;
MCP.Initialize;
MCP.ListPrompts;
MCP.ListResources;
MCP.ListTools;
MCP.RequestTool('GetTemperature', '{"city":"Madrid"}');
end;Let an AI model query your application's data through controlled, typed tools. Read-only or read-write, your handler decides per call.
Expose CRM, ERP or ticketing operations as MCP tools so natural-language instructions drive existing Delphi business logic.
Give assistants like Claude Desktop access to a legacy application's functionality by wrapping it in a stdio MCP server.
Chain tools and resources for retrieval, processing and action in a single AI-driven pipeline, orchestrated from the Delphi side.
What the Model Context Protocol is and how the server and client components divide the work.
Full reference for TsgcWSAPIServer_MCP: properties, methods, events.
Full reference for TsgcWSAPIClient_MCP: catalogue methods and response events.
Step-by-step walkthrough of a working Delphi MCP server.
Connecting a Delphi client to MCP servers and calling their tools.
Securing MCP endpoints: API keys and authenticated flows from Delphi.