Everything sgcAI does, mapped across the 15 components on the SGC AI palette page. Chat with seven model providers through one component, host or consume a Model Context Protocol server, index your own data as vectors, and drive the whole thing by voice. Every license ships full source code. Click a component for its own page with properties, events and examples.
sgcAI is self-contained. Everything in the tables below ships in one package, together with the sgcWebSockets Core runtime the components are built on.
Chat and MCP run on every platform. TsgcAIChat and the two Model Context Protocol components compile for Windows, macOS, Linux, iOS and Android, and the IDE offers them on forms for every platform.
Four components stay Windows only because they call Windows audio and speech APIs: TsgcAudioRecorderMCI, TsgcAudioRecorderWave, TsgcAudioPlayerMCI and TsgcTextToSpeechSystem. See Platforms & languages below.
7 providers, 4 components
Client and server
RAG over your own data
System, Google, Amazon
MCI and Wave
Chat and MCP on Windows, macOS, Linux, iOS and Android
All of them register on a single palette page named SGC AI.
| Component | Class | Family | Description |
|---|---|---|---|
| AI Chat | TsgcAIChat | Chat | One chat client for OpenAI, Anthropic, Gemini, DeepSeek, Ollama, Grok and Mistral. |
| AI Agent | TsgcAIAgent | Chat | Runs the tool loop on every provider with your tools, the tools of your MCP servers and approval for tools that change data. |
| Guardrails | TsgcAIGuardrails | Chat | PII redaction, allowed and denied tools, tool approval and JSON schema validation of answers. |
| Usage Meter | TsgcAIUsageMeter | Chat | Costs per provider and model, budgets per user and session, CSV, TDataSet and OpenTelemetry export. |
| OpenAI ChatBot | TsgcAIOpenAIChatBot | Chat | Voice chatbot: record, transcribe with Whisper, answer, speak the reply. |
| OpenAI Translator | TsgcAIOpenAITranslator | Chat | Speech in, speech out translation into any language: transcription plus chat translation. |
| OpenAI Assistant | TsgcAIOpenAIAssistant | Chat | Assistants API: threads, runs, File Search, Code Interpreter, function calling. |
| MCP Server | TsgcWSAPIServer_MCP | MCP | Publish tools, prompts and resources over HTTP or HTTP Streamable. |
| MCP Client | TsgcWSAPIClient_MCP | MCP | Consume any MCP server over HTTP, HTTP Streamable or stdio. |
| OpenAI Embeddings | TsgcAIOpenAIEmbeddings | Vectors | Generate embeddings from text or a file and query the nearest match. |
| Vector Database (File) | TsgcAIDatabaseVectorFile | Vectors | Local file-backed vector store, no server and no account. |
| Vector Database (Pinecone) | TsgcAIDatabaseVectorPinecone | Vectors | Managed vector index through the Pinecone REST API. |
| Text to Speech (System) | TsgcTextToSpeechSystem | Speech | Speaks through the engine installed on the machine, no key required. |
| Text to Speech (Google) | TsgcTextToSpeechGoogle | Speech | Google Cloud Text-to-Speech with service account authentication. |
| Text to Speech (Amazon) | TsgcTextToSpeechAmazon | Speech | Amazon Polly with signed AWS requests and SSML support. |
| Audio Recorder (MCI) | TsgcAudioRecorderMCI | Audio | Microphone capture to a WAV file with silence detection. |
| Audio Recorder (Wave) | TsgcAudioRecorderWave | Audio | Streaming microphone capture with full control of the sample format. |
| Audio Player (MCI) | TsgcAudioPlayerMCI | Audio | Playback with a genuine end-of-playback notification. |
Seven providers reached through one Provider property, plus three OpenAI components for voice and agent scenarios.
| Capability | Component | Notes |
|---|---|---|
| OpenAI | TsgcAIChat | Provider := aicpOpenAI. GPT models through the Chat Completions API. |
| Anthropic | TsgcAIChat | Provider := aicpAnthropic. Claude models through the Messages API. |
| Google Gemini | TsgcAIChat | Provider := aicpGemini. |
| DeepSeek | TsgcAIChat | Provider := aicpDeepSeek. deepseek-flash and deepseek-v4-pro. |
| Ollama (local models) | TsgcAIChat | Provider := aicpOllama. ChatOptions.BaseUrl points at your own host, nothing leaves the machine. |
| xAI Grok | TsgcAIChat | Provider := aicpGrok. |
| Mistral | TsgcAIChat | Provider := aicpMistral. |
| Blocking call | TsgcAIChat | Chat(aMessage) returns the assistant text. |
| Streaming | TsgcAIChat | ChatStream plus OnChatStream, which carries a Cancel flag to abort mid-generation. |
| Asynchronous | TsgcAIChat | ChatAsync returns a future on Delphi 2010 and later. |
| System prompt | TsgcAIChat | SystemMessage for every turn, or ChatWithSystem for a single turn. |
| Bounded history | TsgcAIChat | MaxHistoryMessages caps the replayed context. ClearHistory and GetHistory manage the transcript. |
| Sampling controls | TsgcAIChat | MaxTokens, Temperature and TopP. Leave the last two unset and the provider default applies. |
| Documents and attachments | TsgcAIChat | ChatWithDocument, ChatWithDocumentBase64, ChatWithAttachments and ChatJSONWithAttachments send PDFs and images with the message, one or several held in a TsgcAIChatAttachments list. |
| Tool calling | TsgcAIChat | ChatWithTools and AddToolResult on all seven providers, each in its own format: OpenAI style tools, Anthropic tool use and Gemini function declarations. |
| AI agent | TsgcAIAgent | Run offers your tools, methods of your objects (AddMethod, Delphi 2010 and later) and the tools of MCPClients, runs the calls, asks OnToolApproval for tools that are not ReadOnly and stops on the answer, Options.MaxSteps, the token limit (aasrMaxTokens), Cancel or an error. OnStep and OnUsage report the run. |
| Model aliases | TsgcAIChat | ModelAliases maps names such as fast or smart to a model or to provider:model, loaded from a file, JSON or a URL. ProviderApiKeys holds the key of each provider. |
| Provider fallback | TsgcAIChat | Fallbacks moves a request to the next provider and model on a connection error, HTTP 5xx, HTTP 429 or an unknown model, with tools and streaming. OnFallback reports every switch. |
| Token usage | TsgcAIChat | LastUsage and OnChatUsage: provider, model, input, output and cached tokens of every request, with UserId and SessionId. LastFinishReason: why the last answer ended, such as stop or length. |
| Costs and budgets | TsgcAIUsageMeter | Prices per million tokens, totals per user and session, total, daily or monthly budgets that can refuse a request, SaveToCSV, ExportToDataSet and OpenTelemetry spans with the GenAI conventions. |
| Guardrails | TsgcAIGuardrails | Redacts e-mails, phones, cards, IBANs, IP addresses and custom patterns before the request and restores them in the answer, AllowedTools and DeniedTools with wildcards, ToolApproval, and JSON schema validation of structured answers. |
| Document extraction | TsgcAIDocumentExtractor | PDF or image plus a JSON schema in, validated JSON out. Presets dpInvoice, dpReceipt and dpIDCard, a confidence per field with OnLowConfidence below MinConfidence, and one automatic retry when the answer does not validate. Link TsgcAIOCRTesseract or TsgcAIOCRWindows, the OCR built into Windows 10 and later, to recognise the text locally, so a text only model on Ollama can extract too. |
| Realtime voice agent | TsgcAIVoiceAgent | Listens, detects turns, talks back and calls tools over OpenAI Realtime or Gemini Live, on WebSocket or, with OpenAI, WebRTC, on Windows, Android and iOS. Server VAD, semantic VAD, client VAD or manual turns, barge-in that stops the agent and truncates its answer, live user and agent transcripts on OnTranscript, tools from OnToolCall and from an assigned MCPClient, and G.711 mu-law or A-law at 8 kHz for telephony. |
| Realtime voice translator | TsgcAIVoiceTranslator | Streams the microphone to OpenAI gpt-realtime-translate and plays the translated speech as it arrives, with live transcripts of the source and of the translation on OnInputTranscript and OnOutputTranscript. Change Language while the session runs, or turn the devices off and push your own PCM16 audio with SendAudio. Windows audio, or your own audio on Linux and macOS. |
| Speech to text | TsgcAIOpenAIChatBot, TsgcAIOpenAITranslator | gpt-transcribe transcription, and translation into any language through a chat model, with an event to review the text before use. |
| Retrieval augmented answers | TsgcAIOpenAIChatBot | Link an Embeddings component and the retrieved context is added before the model is called. |
| Assistants, threads and runs | TsgcAIOpenAIAssistant | CreateAssistant, CreateThread, CreateMessageText, CreateRun, CreateRunAndWait. |
| File Search and Code Interpreter | TsgcAIOpenAIAssistant | AssistantOptions.Tools, with UploadVectorStoreFile to add your own documents. |
| Function calling | TsgcAIOpenAIAssistant | Tools.Functions plus the OnFunctionCall event. |
Both ends of the protocol that AI clients use to reach tools, prompts and resources. MCP 2026-07-28 is supported with full backward compatibility for 2025-11-25.
| Capability | Client | Server | Notes |
|---|---|---|---|
| HTTP JSON-RPC transport | aimcptrHttp | TransportOptions.Http | One request, one response, connection closes. |
| HTTP Streamable transport | aimcptrHttpStreamable | TransportOptions.HttpStreamable | Keep-alive plus an SSE stream, which is what carries notifications and server-initiated requests. |
| stdio transport | aimcptrStdio | separate stdio host | The client spawns a local subprocess. Server-side stdio is provided by a dedicated host class rather than the palette component. |
| Handshake | Initialize | OnMCPInitialize | Capability negotiation and session id assignment. |
| Sessions | Session, HeartBeat | OnMCPSessionNew, OnMCPSessionEnd, SessionTimeout | Idle sessions expire on the server, the client can keep one alive with periodic pings. |
| Tools | ListTools, RequestTool | Tools, OnMCPRequestTool | List and call. The server raises an event where your code does the work. |
| Tool pinning | MCPOptions.ToolPinning, OnMCPToolChanged | — | A SHA-256 fingerprint of every listed tool. A tool whose description, input schema or annotations change is hidden and refused until you accept the change, which stops tool poisoning. |
| Prompts | ListPrompts, RequestPrompt | Prompts, OnMCPRequestPrompt | Same pattern as tools. |
| Resources | ListResources, RequestResource, ListResourceTemplates | Resources, ResourceTemplates, OnMCPRequestResource | Lists accept a cursor for paging. |
| Resource subscriptions | SubscribeResource, UnsubscribeResource | SendNotificationResourcesUpdated | Follow a resource and be told when it changes. |
| List-changed notifications | received | SendNotificationToolsListChanged and the prompt and resource equivalents | Requires the streamable transport. |
| Roots | OnMCPListRoots | RequestRootsList | The server asks the client which filesystem roots it may use. |
| Sampling | OnMCPSamplingCreateMessage | RequestSamplingCreateMessage | The server asks the client's model for a completion. |
| Elicitation | OnMCPElicitationCreate | RequestElicitationCreate | The server asks the user for input through the client. |
| Completion | Complete | handled | Argument value completion for tools and prompts. |
| Logging | SetLoggingLevel | SendLogMessage | Structured log messages at a client-controlled level. |
| Authentication | API key, custom header, OAuth 2.0 | API key, custom header, OAuth 2.0 | MCPOptions.AuthenticationOptions on both sides. |
| Built-in authorization server | — | ✓ | Dynamic client registration, consent page, PKCE and refresh tokens, following RFC 8414 and RFC 9728. |
| Origin validation | — | HttpStreamable.ValidateOrigin | Pins a session to the Origin header it was created with. |
| MCP 2026-07-28 (stateless) | MCPOptions.ProtocolEra, Discover | era detected per request | Per-request _meta, server/discover, validated Mcp-Method, Mcp-Name and Mcp-Param headers. The client falls back to 2025-11-25 automatically. |
| Cache hints | received | MCPOptions.Cache | ttlMs and cacheScope on 2026-07-28 results. |
| subscriptions/listen | SubscriptionsListen, SubscriptionsCancel | MCPOptions.Subscriptions | One stream for list changes and resource updates. |
| Multi Round-Trip Requests | OnMCPInputRequired | aResponse.InputRequired, MCPOptions.MRTR | Elicitation (form and URL), sampling and roots with an HMAC-signed requestState. |
| Tasks extension | MCPOptions.Tasks.Enabled, TasksGet | CreateTask, MCPOptions.Tasks | io.modelcontextprotocol/tasks for long running tool calls; the client polls automatically with AutoPoll. |
| Client authorization | MCPOptions.Authorization | — | Protected resource metadata discovery, Client ID Metadata Documents, dynamic registration fallback, PKCE, resource indicators, RFC 9207 iss validation, per-issuer credentials. |
Turn your own documents into vectors, store them locally or in a managed index, and retrieve context for every question.
| Capability | File store | Pinecone | Notes |
|---|---|---|---|
| Generate embeddings | ✓ | ✓ | TsgcAIOpenAIEmbeddings.CreateEmbeddings and CreateEmbeddingsFromFile. |
| Chunking | ✓ | ✓ | EmbeddingsOptions.ChunkSize, default 4096 characters. |
| Embedding model | ✓ | ✓ | EmbeddingsOptions.Model, default text-embedding-3-small. |
| Batch upsert | write on close | single REST upsert | BeginAddData, AddData, EndAddData on both stores. |
| Nearest neighbour query | QueryData | QueryData | Returns the stored source text behind the closest vector. |
| Retrieve context for a prompt | ✓ | ✓ | GetEmbedding(aInput, aParams) embeds the query and returns the best match. |
| Progress and cancel | ✓ | ✓ | OnCreateEmbeddingsStart, OnCreateEmbeddingsProgress with a continue flag, OnCreateEmbeddingsStop. |
| Per-chunk hooks | ✓ | ✓ | OnBeforeCreateEmbedding can veto a request, OnAfterCreateEmbedding reports request and response. |
| Works offline | ✓ | — | The file store needs no network on query. Generating embeddings still calls the provider. |
| Shared across machines | — | ✓ | A Pinecone index can be written by one application and read by another. |
| Configuration | VectorFileOptions.InputFilename, VectorFilename | PineconeOptions, PineconeIndexOptions | Two files, or an API key plus index and project id. |
| Id control | OnBeginAddData | OnBeginAddData | Set the first id and the id label so several ingests can share a store. |
All three share one interface, so the chatbot, translator and assistant accept any of them.
| Capability | System | Amazon | |
|---|---|---|---|
| Speak a string | TextToSpeech | TextToSpeech | TextToSpeech |
| Needs credentials | no | service account JSON | AWS access key and secret |
| Needs a network | no | yes | yes |
| Cost per character | none | provider pricing | provider pricing |
| Voice selection | machine speech settings | Language, VoiceId, Gender | Engine, VoiceId |
| Output format | direct playback | AudioEncoding, FileName | OutputFormat, FileName |
| SSML | — | — | TextType |
| Audio player required | no | yes, AudioPlayer | yes, AudioPlayer |
| Authentication handling | — | signed JWT with a cached access token | AWS Signature Version 4 |
| Start and stop events | ✓ | ✓ | ✓ |
Two recorders with different trade-offs, and the player that closes the voice loop.
| Capability | Recorder (MCI) | Recorder (Wave) | Player (MCI) |
|---|---|---|---|
| Windows API used | MCI | waveIn | MCI |
| Output | WAV file | live stream | plays a file |
| Sample format control | fixed, 16 bit mono 22050 Hz | SampleRate, Channels, BitsPerSample | — |
| Silence detection | MCIOptions.LevelMin, StopAfter | WaveOptions.LevelMin, StopAfter, WaitForAudio | — |
| Manual or automatic | RecorderOptions.Mode: camoManual or camoAuto | Start and Stop | Play |
| Concurrent instances | ✓, unique MCI alias per component | ✓ | ✓ |
| End of playback notification | — | — | ✓, real MCI notification, not a timer |
| Used by the voice components | default recorder | streaming scenarios | output of the cloud speech components |
What is supported, stated precisely.
| Area | Detail |
|---|---|
| Operating system | Windows, macOS, Linux, iOS and Android. SGC_AI is defined on every platform, so the AI units compile for Win32, Win64, macOS (Intel and ARM), Linux64, iOS and Android. The units that call Windows audio and speech APIs are compiled on Windows only. HTTPS calls need the usual TLS setup of each platform, for example the OpenSSL libraries on Linux. |
| Delphi | Delphi 7 through RAD Studio 13 Florence, VCL and FireMonkey, targeting Win32 and Win64. |
| C++ Builder | C++ Builder 10.1 Berlin through C++ Builder 13, through the generated headers, targeting Win32 and Win64. |
| AI palette components | TsgcAIChat is declared [ComponentPlatforms(0)], so the IDE offers it on every platform. The chatbot, translator, assistant, embeddings, vector, document extractor, OCR and cloud speech components are declared [ComponentPlatforms(pidWin32 or pidWin64)]: the IDE offers them on Win32 and Win64 forms, and their units compile for every platform. TsgcAudioRecorderMCI, TsgcAudioRecorderWave, TsgcAudioPlayerMCI and TsgcTextToSpeechSystem call the Windows MCI, Wave and SAPI APIs and exist on Windows only. |
| MCP components | TsgcWSAPIServer_MCP and TsgcWSAPIClient_MCP are declared [ComponentPlatforms(0)], so they carry no platform restriction and run wherever the sgc HTTP stack runs, including Windows, macOS, Linux, iOS and Android. |
| Bundled runtime | The sgcWebSockets Core runtime is included, so the HTTP client stack, the JSON layer and the HTTP server behind the MCP server API need no separate license. |
| TLS | OpenSSL or SChannel for every provider call, following the sgcWebSockets configuration. |
| Source code | Full Object Pascal source in every paid tier. |
| Redistribution | Binaries you build are royalty free, with no per-seat or per-server runtime fee. |