Copilot/Ollama SDK for C++ and PHP

Github released the Copilot SDK and here 's a C++ wrapper around it to be used in Windows. This also allows to use a local LLama-based model through a local llama-server. I'm already using it in Turbo Play, TurboIRC, FaustChat and other projects.

This is the new RAW only readme. RAW mode does not require python installed. If you want to see the old readme, open README_old.md. Raw.hpp also compiles in linux with g++.

Ollama

If you want to also use local models, download and run Ollama, specify to make it visible to the network. The models will be visible to your C++ code.

Initialize

#include "raw.hpp"

// Use either a running copilot server 
//	COPILOT_RAW raw("127.0.0.1", 3000, true);

// Or a new one with your github access token
//	COPILOT_RAW raw(L"c:\\copilot\\copilot.exe", 3000, "your_token",1);

Quick example

auto s1 = raw.CreateSession("gpt-4.1");
auto r = raw.One(s1, "Tell me a short joke", 60000); // wait for maximum 1 minute

Example

raw.Ping();
auto models = raw.CopilotModels();
std::vector<std::shared_ptr<COPILOT_SESSION>> sessions;
raw.Sessions(sessions);
	
auto as = raw.AuthStatus();
auto st = raw.Status();
/*
struct COPILOT_SESSION_PARAMETERS
{
	bool Streaming = true;
	std::string reasoning_effort;
	std::string system_message;
	std::vector<std::wstring> skill_dirs;
	std::vector<std::wstring> disabled_skills;
	bool Infinite = true;
	std::function<void(nlohmann::json& j, std::string& resp, bool& free_form,long long cb)> user_ask_function;
	long long ask_function_callback = 0;
};
*/
auto s1 = raw.CreateSession("gpt-4.1", nullptr); // use the default COPILOT_SESSION_PARAMETERS if nullptr is passed
//  Ollama also supported
//	auto s1 = raw.CreateSession("phi:latest");
std::vector<std::wstring> files = { L"f:\\tp2imports\\365.jpg" };
auto m1 = raw.CreateMessage(s1, "What do you see in this image?", 0, 0, 0, &files);
auto m2 = raw.CreateMessage(s1, "Please tell me all numbers from 1 to 100", [&](std::string tok, long long ptr) -> HRESULT {
	std::cout << tok;
	if (brk)
	{
		brk = 0;
		return E_ABORT;
	}
	return S_OK;
	}, [&](std::string tok, long long ptr) -> HRESULT {
		std::cout << tok;
		return S_OK;
		}, 0);
raw.Send(s1, m1);
raw.Send(s1, m2);
raw.Wait(s1, m2, 60000);
raw.Wait(s1, m1, 60000);
raw.DestroySession(s1); // leaves files there so session can be resumed later
raw.DeleteSession(s1); // deletes session

User input

You can pass a function callback to the COPILOT_SESSION_PARAMETERS and this is called whenever the model wants to ask you a question.

COPILOT_SESSION_PARAMETERS spp;
spp.user_ask_function = [&](nlohmann::json& j, std::string& resp, bool& free_form,long long cb) -> void {
	// auto question = j["params"]["question"].get<std::string>();
	// auto is_question_free_form = j["params"]["free_form"].get<bool>();
	// auto choices = j["params"]["choices"].get<std::vector<std::string>>(); // maybe not exists if is_question_free_form is true
	resp = "My name is Michael";
	free_form = true;
	};
auto s1 = raw.CreateSession("gpt-4.1", &spp);
auto r = raw.One(s1, "Ask the user his name", 60000);

Tools

Tools are simply function callbacks. You tell the model "Hey, I know the answer to a specific question, so if you need, call my code".

std::vector<COPILOT_TOOL_PARAMETER> params = { {"city","City","City name","string",true}}; // name title description type required
raw.AddTool("GetWeather", "Get the current weather for a city", "GetWeatherParams", params, [&](
	std::string session_id,
	std::string tool_id,
	std::vector<std::tuple<std::string, std::any>>& parameters)
	{
		nlohmann::json j;
		for (auto& p : parameters)
		{
			std::string name;
			std::any value;
			std::tie(name, value) = p;
			if (name == "city")
			{
				j["city"] = std::any_cast<std::string>(value);
			}
		}
		j["condition"] = "Sunny";
		j["temperature"] = "25C";
		// Or you can return a direct string, say "It is sunny".
		return j.dump();
	});
auto s1 = raw.CreateSession("gpt-4.1");
auto m2 = raw.CreateMessage("What is the weather in Seattle?", [&](std::string tok, long long ptr) -> HRESULT {
	std::cout << tok;
	if (brk)
	{
		brk = 0;
		return E_ABORT;
	}
	return S_OK;
	}, [&](std::string tok, long long ptr) -> HRESULT {
		std::cout << tok;
		return S_OK;
		}, 0);
	raw.Send(s1, m2);
	raw.Wait(s1, m2, 600000);
	std::string str = m2->completed_message->reasoningText.c_str();
	str += "\r\n\r\n";
	str += m2->completed_message->content.c_str();
	MessageBoxA(0, str.c_str(), "Information", 0);

Other functions

• Compact() to compact a session
• Status() contains also the account's premium quota percentages and usage
• SetMode(...) to set the current session mode, COPILOT_RAW_MODE:: INTERACTIVE, PLAN, AUTOPILOT.
• DeleteSession and DestroySession to delete a session with or without deleting the files used in it.
• ResumeSession to resume an existing session (get all resumable sessions with Sessions()).
• SwitchModel to switch model in an existing session.

AICmd

AICmd is a tool build with this SDK. Load it, then /help for commands.

PHP

Just require "cop.php" in your PHP code. You do not need python.

<?php
require_once "cop.php";

$cop = new Copilot("your_token","/usr/local/bin/copilot",8765);
// or
$cop = new Copilot("","",8765); // run with an existing server
$m1 = $cop->Ping();
$m1 = $cop->AuthStatus();
$m1 = $cop->Quota();
$m1 = $cop->Sessions();

$s1 = new COPILOT_SESSION_PARAMETERS();
$s1->Streaming = true;
$s1->system_message = "You are a helpful assistant for testing the copilot cli.";
$session_id = $cop->CreateSession("gpt-4.1",$s1,true);
printf("Session ID: %s\n",$session_id); 
// Send message
$m1 = $cop->Prompt($session_id,"What is the capital of France?",true);
// End session
$x1 = $cop->EndSession($session_id,true);

?>

License

MIT