LLMQore

Qt/C++ library for cloud and local LLM providers, MCP clients and servers, and ACP agents. Streaming deltas arrive as Qt signals on the object's own thread, async results as QFuture, ownership follows QObject parent-child. Links against Core, Network and Concurrent.

What it does

• LLM REST API — cloud and local: Claude, OpenAI Chat Completions, OpenAI Responses, Google AI, Mistral, DeepSeek, Qwen, Ollama, llama.cpp; streaming or buffered, images in, reasoning out, token usage, cancellation
• Tool calling — your BaseTool subclass, tool loop included, gated by QFuture<bool> if you want
• MCP — client and server over stdio, Streamable HTTP or legacy SSE, sharing one tool registry; mcp-bridge CLI puts many upstream servers behind one endpoint
• Conversation — one history, translated into each provider's shape, portable between them
• ACP host — any agent in the JSON registry; ships Claude Code and Codex

LLM REST API

Ask and await the answer:

auto *client = new LLMQore::ClaudeClient(
    "https://api.anthropic.com", apiKey, "claude-sonnet-4-5", this);

client->askOnce("What is Qt?").then(this, [this](const LLMQore::CompletionInfo &result) {
    m_view->setPlainText(result.fullText);
});

Or watch it arrive, which is what you want in a chat panel:

connect(client, &LLMQore::BaseClient::accumulatedReceived,
        this, [this](const LLMQore::RequestID &, const QString &answer) {
    m_view->setPlainText(answer);
});

client->ask("What is Qt?");

accumulatedReceived carries the whole answer so far; chunkReceived carries only the new delta. Both are emitted on the client's own thread, so a direct connection into a widget or model is safe.

→ Quick Start

Tool calling

Expose a function over data the provider cannot see — the open document, the current selection, a local database:

client->tools()->addTool(new SearchCurrentFileTool(client));

conversation.addUser("Where do we handle the timeout?");
client->ask(conversation);

The client drives the loop: the model requests the tool, executeAsync runs, the result is sent back, the model answers. setMaxToolContinuations() bounds it, ten rounds by default. toolStarted and toolResultReady report progress; setExecutionGate() gates each call behind a QFuture<bool>.

→ Quick Start · LLM clients

MCP

Using external tools

Tools from an MCP server enter the same registry and reach the model through the same tool-definition array:

client->tools()->addMcpServer({.name = "filesystem", .command = "npx",
    .arguments = {"-y", "@modelcontextprotocol/server-filesystem", "/home/user"}});

client->tools()->loadMcpServers(QJsonDocument::fromJson(configData).object());

loadMcpServers reads the mcpServers object Claude Desktop uses and returns how many servers it registered:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": ["-y", "@modelcontextprotocol/server-filesystem", "/home/user"]
    }
  }
}

Serving your own tools

McpServer::setToolRegistry takes the same ToolsManager the client uses, so one registration serves both the in-process model and any MCP client that connects:

auto *server = new LLMQore::Mcp::McpServer(
    new LLMQore::Mcp::McpHttpServerTransport({.port = 8080, .path = "/mcp"}, this),
    cfg, this);

server->setToolRegistry(client->tools());
server->start();

The registry is shared, not copied: a tool added later reaches both sides, and the server forwards toolsChanged as notifications/tools/list_changed.