Fifteen palette components that put large language models, the Model Context Protocol, embeddings, vector search and speech directly into your Windows application. One chat component reaches seven providers through a single property. sgcAI is a standalone product, it bundles the sgcWebSockets Core runtime it is built on, so there is no base license to buy.
What sgcAI needs from you, and where it runs. Both answers are short.
Unlike the other eSeGeCe add-ons, it does not sit on top of a base product. Everything the components need is bundled inside sgcAI: the sgcWebSockets Core runtime, the HTTP client stack, the JSON layer and the HTTP server behind the MCP server API. One SKU, one installer, nothing else to buy.
The AI package is compiled under MSWINDOWS only, so 13 of the 15 palette components are registered for Win32 and Win64. They do not build for macOS, Linux, iOS or Android.
The two Model Context Protocol components, TsgcWSAPIServer_MCP and TsgcWSAPIClient_MCP, are the exception. They carry no platform restriction and run wherever the sgc HTTP stack runs, including macOS, Linux, iOS and Android. See the full platform table.
Every component is compiled Object Pascal with design-time properties and events. You set an API key in the Object Inspector and handle a response event. There is no Python bridge, no scripting host and no external process.
TsgcAIChat is the unified client. Set Provider, set ChatOptions.ApiKey and ChatOptions.Model, call Chat. Token streaming arrives on OnChatStream with a Cancel flag, and ChatAsync returns a future on Delphi 2010 and later so the UI thread never blocks. The other three wrap complete conversation loops, including the voice ones.
Publish your own tools, prompts and resources to any MCP-aware client with TsgcWSAPIServer_MCP, including sessions, notifications and server-initiated requests. Consume someone else's server with TsgcWSAPIClient_MCP over HTTP, HTTP Streamable or stdio, authenticating with an API key, a custom header or OAuth2. These two carry no platform restriction, so they compile everywhere the sgc HTTP stack does.
TsgcAIOpenAIEmbeddings turns text or a whole file into vectors, raising progress events across the batch, and writes them to whichever store is assigned to Database. Two backends ship: a local file through TsgcAIDatabaseVectorFile, which needs no server, no account and no network round trip, and a managed index through TsgcAIDatabaseVectorPinecone. Query the nearest neighbours and feed them back to the model as context.
Capture the microphone through the Windows MCI or Wave API, transcribe with Whisper, send the text to a model, then speak the answer back. Three interchangeable speech engines share one interface: the engine already installed on the machine, which needs no API key and no network, plus Google Cloud Text-to-Speech and Amazon Polly. The chatbot, translator and assistant components wire the whole loop for you.
Provider PropertyThese are not seven components. They are the seven values of TsgcAIChatProvider, so changing vendor changes one assignment and nothing else. The same Chat, ChatStream and ChatAsync calls, the same history handling and the same events apply to all of them.
aicpOpenAI
GPT models through the Chat Completions API.
aicpAnthropic
Claude models through the Messages API, with the same streaming and history handling as the rest.
aicpGemini
Gemini models through the Google Generative Language API.
aicpDeepSeek
deepseek-chat and deepseek-reasoner.
aicpOllama
Local models. Point ChatOptions.BaseUrl at your Ollama host and no data leaves the machine.
aicpGrok
Grok models through the xAI API.
aicpMistral
Mistral models through the Mistral API.
Pick a provider, set the key and the model, call Chat. Switching to a different vendor later is one assignment.
uses
sgcAI, sgcAI_Chat;
var
oChat: TsgcAIChat;
vAnswer: string;
begin
oChat := TsgcAIChat.Create(nil);
oChat.Provider := aicpAnthropic;
oChat.ChatOptions.ApiKey := 'sk-ant-...';
oChat.ChatOptions.Model := 'claude-sonnet-4-5';
oChat.SystemMessage := 'You are a Delphi expert. Answer in one paragraph.';
vAnswer := oChat.Chat('How do I keep a WebSocket connection alive?');
// Same code, different vendor: one property and one key.
oChat.Provider := aicpOpenAI;
oChat.ChatOptions.ApiKey := 'sk-...';
oChat.ChatOptions.Model := 'gpt-4o';
vAnswer := oChat.Chat('Same question, different model.');
end;// includes: sgcAI.hpp, sgcAI_Chat.hpp
TsgcAIChat *oChat = new TsgcAIChat(NULL);
oChat->Provider = aicpAnthropic;
oChat->ChatOptions->ApiKey = "sk-ant-...";
oChat->ChatOptions->Model = "claude-sonnet-4-5";
oChat->SystemMessage = "You are a Delphi expert. Answer in one paragraph.";
String vAnswer = oChat->Chat("How do I keep a WebSocket connection alive?");
// Same code, different vendor: one property and one key.
oChat->Provider = aicpOpenAI;
oChat->ChatOptions->ApiKey = "sk-...";
oChat->ChatOptions->Model = "gpt-4o";
vAnswer = oChat->Chat("Same question, different model.");uses
sgcAI, sgcAI_MCP_Classes, sgcAI_MCP_Types;
var
oMCP: TsgcWSAPIClient_MCP;
begin
oMCP := TsgcWSAPIClient_MCP.Create(nil);
// Remote server over streamable HTTP.
oMCP.MCPOptions.Transport := aimcptrHttpStreamable;
oMCP.MCPOptions.HttpOptions.URL := 'https://mcp.example.com/mcp';
oMCP.MCPOptions.ClientInfo.Name := 'sgc-mcp-client';
oMCP.MCPOptions.ClientInfo.Version := '1.0.0';
oMCP.OnMCPListTools := MCPListTools;
oMCP.OnMCPResponseTool := MCPToolResponse;
oMCP.Initialize;
oMCP.ListTools;
oMCP.RequestTool('GetTemperature', vArguments); // vArguments: IsgcJSON
end;
// Or spawn a local MCP server as a subprocess instead:
oMCP.MCPOptions.Transport := aimcptrStdio;
oMCP.MCPOptions.StdioOptions.Command := 'C:\tools\my-mcp-server.exe';
oMCP.MCPOptions.StdioOptions.Arguments := '--stdio';Tokens arrive on OnChatStream with a Cancel flag, so a long generation can be stopped mid-flight.
ChatAsync returns a future on Delphi 2010 and later, so the main thread stays responsive during a model call.
SystemMessage sets the persona once and MaxHistoryMessages caps how much of the transcript is replayed each turn, which keeps token cost predictable.
Requests go straight from your application to the vendor endpoint over the sgc HTTP client. There is no eSeGeCe relay in the path.
The AI package is gated behind MSWINDOWS in sgcVer.inc. That single define is what shapes the table below, so read it before you plan a mobile or Linux build.
TsgcWSAPIServer_MCP and TsgcWSAPIClient_MCP, no platform restriction
sgcAI is licensed on its own, starting at €149 for a single developer. All licenses include full source code, 1 year of updates and a 50% to 70% renewal discount: 50% when you renew one pack, 60% for two, 70% for three or more. sgcAI and sgcMQ each count as a pack.
€149
Single, Team and Site licenses available.
Pair sgcAI with our other Delphi, C++ Builder and .NET component libraries.
Enterprise WebSocket, HTTP/2, MQTT, AMQP and WebRTC components. The Enterprise edition already includes the sgcAI components.
Learn more →QUIC (RFC 9000) and HTTP/3 (RFC 9114) client and server components built on the native OpenSSL 3.5 QUIC engine.
Learn more →Data-aware server-side HTML/UI components for Delphi, C++ Builder and .NET, with Bootstrap 5 and htmx.
Learn more →Enterprise digital signatures. XAdES, PAdES, CAdES and ASiC with 10 key providers and 21 EU country profiles.
Learn more →OpenAPI 3.0 parser and SDK generator. Turn any OpenAPI spec into a strongly-typed Delphi client in seconds.
Learn more →Native Windows Hello, fingerprint and Windows Biometric Framework components for Delphi and C++ Builder.
Learn more →