Custom Agent

Prerequisites: Read the C++ Framework Overview and Integration Guide first. Your project should already link against gaia::gaia_core.

What You Can Customize

gaia::Agent is designed to be subclassed. All domain-specific behavior lives in two virtual methods:

Override	Purpose
`getSystemPrompt()`	Control how the LLM thinks and responds — persona, constraints, output format
`registerTools()`	Register native C++ callbacks and/or connect MCP servers

Two additional hooks are available without subclassing:

API	Purpose
`setOutputHandler()`	Replace the console with a custom output capture
`AgentConfig` fields	Tune step limits, debug verbosity, silent mode

Step 1 — Custom System Prompt

getSystemPrompt() returns a string that is injected into the LLM system message before the auto-generated tool list. Use it to:

Define the agent’s persona and constraints
Specify the output format you expect
Guide reasoning style (chain-of-thought, step-by-step, etc.)

#include <gaia/agent.h>

class WeatherAgent : public gaia::Agent {
public:
    WeatherAgent() : Agent(makeConfig()) { init(); }

protected:
    std::string getSystemPrompt() const override {
        return
            "You are a meteorology assistant. "
            "Always answer in concise bullet points. "
            "Never speculate — if you do not have data, say so explicitly. "
            "When reporting temperatures always use both Celsius and Fahrenheit.";
    }

    // ... registerTools() below ...

private:
    static gaia::AgentConfig makeConfig() {
        gaia::AgentConfig cfg;
        cfg.modelId = "Qwen3-4B-GGUF";
        return cfg;
    }
};

The framework automatically appends the tool list and response-format schema after your system prompt. You do not need to describe tools in getSystemPrompt() — they are injected automatically.

Step 2 — Register Tools with Typed Parameters

registerTools() is called by init() once, at construction time. Each tool has:

A name — used by the LLM to call it
A description — shown in the LLM system prompt
A callback — std::function<json(const json&)>
An optional list of typed parameters — passed as std::vector<gaia::ToolParameter>

Parameter Types

`ToolParamType`	JSON Schema type	C++ usage
`STRING`	`"string"`	Text, identifiers, paths
`INTEGER`	`"integer"`	Counts, indices
`NUMBER`	`"number"`	Floats, measurements
`BOOLEAN`	`"boolean"`	Flags, toggles
`ARRAY`	`"array"`	Lists of values
`OBJECT`	`"object"`	Nested structures

Example: Multiple Tools with Typed Parameters

#include <gaia/agent.h>
#include <gaia/types.h>
#include <stdexcept>

class WeatherAgent : public gaia::Agent {
protected:
    void registerTools() override {

        // Tool 1: no parameters
        toolRegistry().registerTool(
            "get_supported_cities",
            "Return the list of cities for which weather data is available.",
            [](const gaia::json& /*args*/) -> gaia::json {
                return {{"cities", {"Seattle", "Austin", "London", "Tokyo"}}};
            },
            {}  // no parameters
        );

        // Tool 2: one required STRING parameter
        toolRegistry().registerTool(
            "get_current_weather",
            "Return the current weather conditions for a given city.",
            [](const gaia::json& args) -> gaia::json {
                std::string city = args.value("city", "");
                if (city.empty()) throw std::runtime_error("city is required");

                // Replace with real API call in production
                return {
                    {"city",        city},
                    {"temperature", "18°C / 64°F"},
                    {"conditions",  "Partly cloudy"},
                    {"humidity",    "72%"}
                };
            },
            {
                {"city", gaia::ToolParamType::STRING, /*required=*/true,
                 "The city name to get weather for (e.g. 'Seattle')"}
            }
        );

        // Tool 3: mixed required + optional parameters
        toolRegistry().registerTool(
            "get_weather_forecast",
            "Return a multi-day weather forecast for a city.",
            [](const gaia::json& args) -> gaia::json {
                std::string city = args.value("city", "");
                int days        = args.value("days", 3);  // optional, defaults to 3

                if (city.empty()) throw std::runtime_error("city is required");
                if (days < 1 || days > 7) throw std::runtime_error("days must be 1-7");

                gaia::json forecast = gaia::json::array();
                for (int i = 1; i <= days; ++i) {
                    forecast.push_back({
                        {"day",         "Day " + std::to_string(i)},
                        {"high",        "20°C / 68°F"},
                        {"low",         "12°C / 54°F"},
                        {"conditions",  "Mostly sunny"}
                    });
                }
                return {{"city", city}, {"forecast", forecast}};
            },
            {
                {"city", gaia::ToolParamType::STRING,  /*required=*/true,
                 "The city name to forecast"},
                {"days", gaia::ToolParamType::INTEGER, /*required=*/false,
                 "Number of forecast days (1-7, default 3)"}
            }
        );
    }
};

The ToolParameter aggregate is {name, type, required, description}. Optional parameters should have a matching default in your callback (use args.value("key", default_value) from nlohmann/json).

Step 3 — Connect a Custom MCP Server

Call connectMcpServer() after init() to register tools from an external MCP server. This works with any stdio-based MCP server — your own, a third-party package, or a local script.

class WeatherAgent : public gaia::Agent {
public:
    WeatherAgent() : Agent(makeConfig()) {
        init();  // registers native tools first

        // Connect a custom MCP server for extended data sources
        connectMcpServer("weather_data", {
            {"command", "uvx"},
            {"args",    {"my-weather-mcp-server", "--api-key", "YOUR_KEY"}}
        });
    }
    // ...
};

All tools from weather_data are automatically prefixed as mcp_weather_data_<tool_name> and injected into the LLM system prompt alongside your native tools.

Connecting a Local Script

// Python MCP server
connectMcpServer("alerts", {
    {"command", "python"},
    {"args",    {"/path/to/alert_server.py"}}
});

// Node.js MCP server
connectMcpServer("maps", {
    {"command", "node"},
    {"args",    {"/path/to/maps_server.js"}}
});

// Pre-built package via uvx
connectMcpServer("filesystem", {
    {"command", "uvx"},
    {"args",    {"mcp-server-filesystem", "/tmp"}}
});

MCP connections use stdio transport (JSON-RPC 2.0 over stdin/stdout). The server subprocess is spawned at connectMcpServer() time and cleaned up when the agent is destroyed.

Disconnect when Done

agent.disconnectMcpServer("weather_data");   // disconnect one server
agent.disconnectAllMcp();                    // disconnect all servers

Step 4 — Combine Native Tools and MCP Tools

Here is a complete agent that uses both native C++ tools and an MCP server in one class:

combined_agent.cpp

#include <gaia/agent.h>
#include <gaia/types.h>
#include <ctime>
#include <iostream>
#include <stdexcept>
#include <string>

class WeatherAgent : public gaia::Agent {
public:
    WeatherAgent() : Agent(makeConfig()) {
        init();

        // Native tools are already registered by init().
        // Now attach an MCP server for additional data.
        connectMcpServer("weather_data", {
            {"command", "uvx"},
            {"args",    {"my-weather-mcp-server"}}
        });
    }

protected:
    std::string getSystemPrompt() const override {
        return
            "You are a professional meteorology assistant. "
            "Answer in clear, structured bullet points. "
            "Always check current conditions before giving a forecast. "
            "If a city is not supported, say so rather than guessing.";
    }

    void registerTools() override {
        // Native: fast, no subprocess needed
        toolRegistry().registerTool(
            "get_supported_cities",
            "Return the list of cities for which weather data is available.",
            [](const gaia::json&) -> gaia::json {
                return {{"cities", {"Seattle", "Austin", "London", "Tokyo"}}};
            },
            {}
        );

        toolRegistry().registerTool(
            "get_current_weather",
            "Return current weather for a city.",
            [](const gaia::json& args) -> gaia::json {
                std::string city = args.value("city", "");
                if (city.empty()) throw std::runtime_error("city parameter required");
                return {{"city", city}, {"temperature", "18°C / 64°F"}, {"conditions", "Partly cloudy"}};
            },
            {{"city", gaia::ToolParamType::STRING, true, "City name"}}
        );

        // MCP tools from "weather_data" will be added automatically by
        // connectMcpServer() after init() completes.
    }

private:
    static gaia::AgentConfig makeConfig() {
        gaia::AgentConfig cfg;
        cfg.maxSteps = 15;
        cfg.debug    = false;
        return cfg;
    }
};

int main() {
    try {
        WeatherAgent agent;
        auto result = agent.processQuery(
            "What is the current weather in Seattle, and what is the 3-day forecast?"
        );
        std::cout << result["result"].get<std::string>() << "\n";
    } catch (const std::exception& e) {
        std::cerr << "Error: " << e.what() << "\n";
        return 1;
    }
}

Step 5 — Tune AgentConfig

Adjust AgentConfig in your makeConfig() static method to match your use case:

static gaia::AgentConfig makeConfig() {
    gaia::AgentConfig cfg;

    // LLM backend
    cfg.baseUrl = "http://localhost:8000/api/v1";   // default; override for remote server
    cfg.modelId = "Qwen3-4B-GGUF";

    // Loop limits
    cfg.maxSteps              = 15;  // max agent loop iterations
    cfg.maxPlanIterations     = 3;   // max replan cycles before forcing completion
    cfg.maxConsecutiveRepeats = 4;   // repeated identical responses trigger loop detection

    // Debugging
    cfg.debug       = true;   // verbose step-by-step output
    cfg.showPrompts = true;   // print full system prompt and LLM responses

    // Output suppression (see Step 6)
    cfg.silentMode  = false;  // set true to suppress all console output

    return cfg;
}

Field	When to change
`maxSteps`	Increase for complex multi-tool queries; decrease for latency-sensitive paths
`maxPlanIterations`	Increase if the agent frequently replans on large tasks
`debug = true`	During development to trace every LLM call and tool result
`showPrompts = true`	To inspect the exact system prompt sent to the model
`silentMode = true`	For library/service use where you only want the JSON result

Step 6 — Capture Output Programmatically

By default the agent prints to the terminal using TerminalConsole. You can replace it with SilentConsole (built-in) or a fully custom OutputHandler subclass.

Silent Mode (JSON result only)

// Option A: set in config (suppresses all output including final answer)
cfg.silentMode = true;

// Option B: SilentConsole — suppresses progress but still prints the final answer
agent.setOutputHandler(std::make_unique<gaia::SilentConsole>(/*silenceFinalAnswer=*/false));

// Option C: full silence including the final answer
agent.setOutputHandler(std::make_unique<gaia::SilentConsole>(/*silenceFinalAnswer=*/true));

Custom Output Capture

Override OutputHandler to route output to a log file, a UI widget, or any other sink:

capturing_console.h

#include <gaia/console.h>
#include <string>
#include <vector>

/// Captures all agent output in memory for programmatic inspection.
class CapturingConsole : public gaia::OutputHandler {
public:
    // Implement required pure-virtual methods
    void printProcessingStart(const std::string& query, int, const std::string&) override {
        log("QUERY: " + query);
    }
    void printStepHeader(int step, int limit) override {
        log("--- Step " + std::to_string(step) + "/" + std::to_string(limit) + " ---");
    }
    void printStateInfo(const std::string& msg) override  { log("[state] " + msg); }
    void printThought(const std::string& msg) override    { log("[thought] " + msg); }
    void printGoal(const std::string& msg) override       { log("[goal] " + msg); }
    void printPlan(const gaia::json& plan, int) override  { log("[plan] " + plan.dump()); }
    void printToolUsage(const std::string& name) override { log("[tool] calling " + name); }
    void printToolComplete() override                     { log("[tool] done"); }
    void prettyPrintJson(const gaia::json& data, const std::string& title) override {
        log("[json:" + title + "] " + data.dump(2));
    }
    void printError(const std::string& msg) override   { log("[error] " + msg); }
    void printWarning(const std::string& msg) override { log("[warn] " + msg); }
    void printInfo(const std::string& msg) override    { log("[info] " + msg); }
    void startProgress(const std::string& msg) override { log("[progress] " + msg); }
    void stopProgress() override                        { log("[progress] done"); }
    void printFinalAnswer(const std::string& answer) override {
        log("[answer] " + answer);
        finalAnswer_ = answer;
    }
    void printCompletion(int steps, int limit) override {
        log("[done] " + std::to_string(steps) + "/" + std::to_string(limit) + " steps");
    }

    // Accessors
    const std::vector<std::string>& lines() const { return lines_; }
    const std::string& finalAnswer() const         { return finalAnswer_; }

private:
    void log(const std::string& line) { lines_.push_back(line); }
    std::vector<std::string> lines_;
    std::string finalAnswer_;
};

Attach it to your agent after construction:

auto console = std::make_unique<CapturingConsole>();
CapturingConsole* consolePtr = console.get();  // keep raw pointer before move

agent.setOutputHandler(std::move(console));

auto result = agent.processQuery("What is the weather in Seattle?");

// Inspect all captured lines
for (const auto& line : consolePtr->lines()) {
    std::cout << line << "\n";
}

// Or just the final answer
std::cout << consolePtr->finalAnswer() << "\n";

Step 7 — Embedding in Your Application

When integrating the agent into a desktop application (WPF, Qt, Electron), you need to run it headless, capture its output programmatically, and keep your UI responsive.

Headless Pattern

Use SilentConsole to suppress all terminal output. The only interface is the JSON return value from processQuery():

#include <gaia/agent.h>
#include <gaia/console.h>

gaia::AgentConfig cfg;
cfg.silentMode = true;  // uses SilentConsole automatically

MyAgent agent;
// No terminal output — agent runs silently
gaia::json result = agent.processQuery("diagnose network issues");

std::string answer = result["result"].get<std::string>();
int steps = result["steps_taken"].get<int>();

Background Thread Pattern

Since processQuery() blocks, run it on a background thread and post results to your UI:

#include <thread>
#include <future>

// Launch agent on background thread
std::future<gaia::json> future = std::async(std::launch::async, [&agent]() {
    return agent.processQuery("run full diagnostic");
});

// UI thread can check if done or wait
if (future.wait_for(std::chrono::seconds(0)) == std::future_status::ready) {
    gaia::json result = future.get();
    // Update UI with result
}

Custom OutputHandler for UI Integration

For real-time progress updates in a GUI, implement a custom OutputHandler that forwards events to your UI framework:

class UIOutputHandler : public gaia::OutputHandler {
public:
    // Callback types — set these to your UI update functions
    std::function<void(int step, int total)> onStepStart;
    std::function<void(const std::string&)>  onThought;
    std::function<void(const std::string&)>  onToolCall;
    std::function<void(const gaia::json&)>   onToolResult;
    std::function<void(const std::string&)>  onAnswer;
    std::function<void(const std::string&)>  onError;

    void printStepHeader(int step, int limit) override {
        if (onStepStart) onStepStart(step, limit);
    }
    void printThought(const std::string& t) override {
        if (onThought) onThought(t);
    }
    void printToolUsage(const std::string& name) override {
        if (onToolCall) onToolCall(name);
    }
    void prettyPrintJson(const gaia::json& data, const std::string&) override {
        if (onToolResult) onToolResult(data);
    }
    void printFinalAnswer(const std::string& answer) override {
        if (onAnswer) onAnswer(answer);
    }
    void printError(const std::string& msg) override {
        if (onError) onError(msg);
    }

    // Minimal no-op implementations for remaining required methods
    void printProcessingStart(const std::string&, int, const std::string&) override {}
    void printStateInfo(const std::string&) override {}
    void printGoal(const std::string&) override {}
    void printPlan(const gaia::json&, int) override {}
    void printToolComplete() override {}
    void printWarning(const std::string&) override {}
    void printInfo(const std::string&) override {}
    void startProgress(const std::string&) override {}
    void stopProgress() override {}
    void printCompletion(int, int) override {}
};

Wire it up:

auto ui = std::make_unique<UIOutputHandler>();
ui->onStepStart = [](int step, int total) {
    // Update progress bar: step / total
};
ui->onThought = [](const std::string& thought) {
    // Display agent reasoning in a text panel
};
ui->onAnswer = [](const std::string& answer) {
    // Show final answer in result area
};

agent.setOutputHandler(std::move(ui));

Thread safety: The OutputHandler methods are called from the thread running processQuery(). If your callbacks update a GUI, you must post to the UI thread (e.g., QMetaObject::invokeMethod in Qt, Dispatcher.Invoke in WPF, PostMessage in Win32).

Multiple Agents

Each Agent instance is independent. You can run multiple agents concurrently on separate threads — each with its own tools, MCP connections, and output handler:

// Each agent gets its own config, tools, and output handler
MyDiagnosticAgent diagAgent;
MyReportAgent reportAgent;

// Safe to run in parallel — no shared state
auto f1 = std::async(std::launch::async, [&] { return diagAgent.processQuery("check health"); });
auto f2 = std::async(std::launch::async, [&] { return reportAgent.processQuery("generate report"); });

Complete Working Example

Below is a self-contained agent combining all the customization points above. Copy it as a starting point for your own agent:

my_custom_agent.cpp

#include <gaia/agent.h>
#include <gaia/types.h>
#include <gaia/console.h>

#include <cstdlib>
#include <iostream>
#include <stdexcept>
#include <string>

#ifdef _WIN32
#  include <windows.h>
#else
#  include <unistd.h>
#endif

// --- Step 6: custom output capture ---
class LoggingConsole : public gaia::OutputHandler {
public:
    void printProcessingStart(const std::string& q, int, const std::string&) override { log("query: " + q); }
    void printStepHeader(int s, int l) override { log("step " + std::to_string(s) + "/" + std::to_string(l)); }
    void printStateInfo(const std::string& m)  override { log("state: " + m); }
    void printThought(const std::string& m)    override { log("thought: " + m); }
    void printGoal(const std::string& m)       override { log("goal: " + m); }
    void printPlan(const gaia::json& p, int)   override { log("plan: " + p.dump()); }
    void printToolUsage(const std::string& n)  override { log("call: " + n); }
    void printToolComplete()                   override { log("call: done"); }
    void prettyPrintJson(const gaia::json& d, const std::string& t) override { log(t + ": " + d.dump()); }
    void printError(const std::string& m)      override { log("ERROR: " + m); }
    void printWarning(const std::string& m)    override { log("WARN: " + m); }
    void printInfo(const std::string& m)       override { log("info: " + m); }
    void startProgress(const std::string& m)   override { log("progress: " + m); }
    void stopProgress()                        override { log("progress: done"); }
    void printFinalAnswer(const std::string& a) override {
        std::cout << "\n=== Answer ===\n" << a << "\n";
    }
    void printCompletion(int s, int l) override {
        std::cout << "[done: " << s << "/" << l << " steps]\n";
    }
private:
    void log(const std::string& m) { std::cout << "[agent] " << m << "\n"; }
};

// --- The custom agent ---
class MyCustomAgent : public gaia::Agent {
public:
    MyCustomAgent() : Agent(makeConfig()) {
        // Step 6: attach custom console before init()
        setOutputHandler(std::make_unique<LoggingConsole>());

        // Steps 1-3 happen inside init()
        init();

        // Step 3: connect MCP server AFTER init()
        connectMcpServer("system_info", {
            {"command", "uvx"},
            {"args",    {"mcp-server-system-info"}}
        });
    }

protected:
    // Step 1: custom persona and constraints
    std::string getSystemPrompt() const override {
        return
            "You are a helpful system information assistant. "
            "Provide concise, accurate answers using only the data from your tools. "
            "If you cannot find data for a request, say so explicitly.";
    }

    // Step 2: native tools with typed parameters
    void registerTools() override {
        // Tool with no parameters
        toolRegistry().registerTool(
            "get_hostname",
            "Return the hostname of the current machine.",
            [](const gaia::json&) -> gaia::json {
                char buf[256] = {};
#ifdef _WIN32
                DWORD sz = sizeof(buf);
                GetComputerNameA(buf, &sz);
#else
                gethostname(buf, sizeof(buf));
#endif
                return {{"hostname", std::string(buf)}};
            },
            {}
        );

        // Tool with a required STRING parameter
        toolRegistry().registerTool(
            "get_env_variable",
            "Return the value of an environment variable.",
            [](const gaia::json& args) -> gaia::json {
                std::string name = args.value("name", "");
                if (name.empty()) throw std::runtime_error("name is required");
                const char* val = std::getenv(name.c_str());
                if (!val) return {{"name", name}, {"value", nullptr}, {"found", false}};
                return {{"name", name}, {"value", std::string(val)}, {"found", true}};
            },
            {
                {"name", gaia::ToolParamType::STRING, /*required=*/true,
                 "The environment variable name (e.g. 'PATH', 'HOME')"}
            }
        );

        // Tool with required + optional parameters
        toolRegistry().registerTool(
            "repeat_message",
            "Repeat a message a given number of times.",
            [](const gaia::json& args) -> gaia::json {
                std::string msg = args.value("message", "");
                int count       = args.value("count", 1);
                if (msg.empty()) throw std::runtime_error("message is required");
                gaia::json out = gaia::json::array();
                for (int i = 0; i < count; ++i) out.push_back(msg);
                return {{"repetitions", out}};
            },
            {
                {"message", gaia::ToolParamType::STRING,  /*required=*/true,  "The message to repeat"},
                {"count",   gaia::ToolParamType::INTEGER, /*required=*/false, "How many times (default 1)"}
            }
        );
    }

private:
    // Step 5: tune AgentConfig
    static gaia::AgentConfig makeConfig() {
        gaia::AgentConfig cfg;
        cfg.baseUrl              = "http://localhost:8000/api/v1";
        cfg.modelId              = "Qwen3-Coder-30B-A3B-Instruct-GGUF";
        cfg.maxSteps             = 15;
        cfg.maxPlanIterations    = 3;
        cfg.maxConsecutiveRepeats = 4;
        cfg.debug                = false;  // set true during development
        cfg.silentMode           = false;  // false: use our LoggingConsole
        return cfg;
    }
};

int main() {
    try {
        MyCustomAgent agent;
        auto result = agent.processQuery("What is the machine hostname and the value of the PATH variable?");
        // Final answer was already printed by LoggingConsole::printFinalAnswer()
        (void)result;
    } catch (const std::exception& e) {
        std::cerr << "Error: " << e.what() << "\n";
        return 1;
    }
}

CMakeLists.txt

cmake_minimum_required(VERSION 3.14)
project(my_custom_agent LANGUAGES CXX)
set(CMAKE_CXX_STANDARD 17)

include(FetchContent)
FetchContent_Declare(
    gaia_core
    GIT_REPOSITORY https://github.com/amd/gaia.git
    GIT_TAG        main
    SOURCE_SUBDIR  cpp
)
FetchContent_MakeAvailable(gaia_core)

add_executable(my_custom_agent my_custom_agent.cpp)
target_link_libraries(my_custom_agent PRIVATE gaia::gaia_core)

Summary

Customization	Where	Notes
Persona, constraints, output format	`getSystemPrompt()`	Injected before tool list
Native C++ tools	`registerTools()` via `toolRegistry().registerTool()`	Typed params improve LLM tool-calling accuracy
External MCP server tools	`connectMcpServer()` after `init()`	Prefixed as `mcp_<server>_<tool>`
Step limits, replanning, debug verbosity	`AgentConfig` in `makeConfig()`	Tune per use-case
Console output routing	`setOutputHandler()`	Call before or after `init()`

Next Steps

C++ Framework Overview

AgentConfig reference and full project structure

Integration Guide

FetchContent, find_package, and shared library integration methods

MCP Client

How the MCP stdio transport works under the hood

Wi-Fi Troubleshooter Agent

Full network diagnostic and auto-fix using registered C++ tools

C++ Source Code

Browse the full implementation on GitHub

Getting Started

C++ Framework

Python Framework

What You Can Customize

Step 1 — Custom System Prompt

Step 2 — Register Tools with Typed Parameters

Parameter Types

Example: Multiple Tools with Typed Parameters

Step 3 — Connect a Custom MCP Server

Connecting a Local Script

Disconnect when Done

Step 4 — Combine Native Tools and MCP Tools

Step 5 — Tune AgentConfig

Step 6 — Capture Output Programmatically

Silent Mode (JSON result only)

Custom Output Capture

Step 7 — Embedding in Your Application

Headless Pattern

Background Thread Pattern

Custom OutputHandler for UI Integration

Multiple Agents

Complete Working Example

Summary

Next Steps

C++ Framework Overview

Integration Guide

MCP Client

Wi-Fi Troubleshooter Agent

C++ Source Code

Getting Started

C++ Framework

Python Framework

​What You Can Customize

​Step 1 — Custom System Prompt

​Step 2 — Register Tools with Typed Parameters

​Parameter Types

​Example: Multiple Tools with Typed Parameters

​Step 3 — Connect a Custom MCP Server

​Connecting a Local Script

​Disconnect when Done

​Step 4 — Combine Native Tools and MCP Tools

​Step 5 — Tune AgentConfig

​Step 6 — Capture Output Programmatically

​Silent Mode (JSON result only)

​Custom Output Capture

​Step 7 — Embedding in Your Application

​Headless Pattern

​Background Thread Pattern

​Custom OutputHandler for UI Integration

​Multiple Agents

​Complete Working Example

​Summary

​Next Steps

C++ Framework Overview

Integration Guide

MCP Client

Wi-Fi Troubleshooter Agent

C++ Source Code

What You Can Customize

Step 1 — Custom System Prompt

Step 2 — Register Tools with Typed Parameters

Parameter Types

Example: Multiple Tools with Typed Parameters

Step 3 — Connect a Custom MCP Server

Connecting a Local Script

Disconnect when Done

Step 4 — Combine Native Tools and MCP Tools

Step 5 — Tune AgentConfig

Step 6 — Capture Output Programmatically

Silent Mode (JSON result only)

Custom Output Capture

Step 7 — Embedding in Your Application

Headless Pattern

Background Thread Pattern

Custom OutputHandler for UI Integration

Multiple Agents

Complete Working Example

Summary

Next Steps