Providers

OpenAI (ChatGPT native transport)

• Uses ChatGPT OAuth login and stores tokens under ~/.intelligencex/auth.json
• Best for users with an existing ChatGPT subscription
• Works with the EasySession and IntelligenceXClient helpers
• Supports native tool calling via ToolRunner and ToolRegistry

using IntelligenceX.OpenAI;

var session = await EasySession.StartAsync();
var result = await session.AskAsync("Summarize the latest PR");
Console.WriteLine(result.Text);

OpenAI (app-server / Codex)

• Launches the Codex app-server process locally
• Best when you want full app-server features (threads, tools, MCP)
• Works via AppServerClient or IntelligenceXClient with AppServer transport

using IntelligenceX.OpenAI.AppServer;

await using var client = await AppServerClient.StartAsync();
await client.InitializeAsync(new ClientInfo("IntelligenceX", "Demo", "1.0"));
var thread = await client.StartThreadAsync("gpt-5.3-codex");
await client.StartTurnAsync(thread.Id, "Hello from app-server");

OpenAI-compatible HTTP (local providers)

• Uses an OpenAI-compatible HTTP endpoint (OpenAI Chat Completions-style)
• Useful for local/self-hosted model servers (Ollama, LM Studio, llama.cpp server, etc.)
• Does not require ChatGPT OAuth

using IntelligenceX.OpenAI;

var options = new IntelligenceXClientOptions {
    TransportKind = OpenAITransportKind.CompatibleHttp,
    DefaultModel = "llama3.1"
};
options.CompatibleHttpOptions.BaseUrl = "http://127.0.0.1:11434";
options.CompatibleHttpOptions.AllowInsecureHttp = true; // loopback http:// opt-in
options.CompatibleHttpOptions.Streaming = true;

await using var client = await IntelligenceXClient.ConnectAsync(options);
var turn = await client.ChatAsync("Summarize the last PR");
Console.WriteLine(EasyChatResult.FromTurn(turn).Text);

Notes:

• BaseUrl may be either http://host:port or http://host:port/v1; it will be normalized to /v1/.
• Inline image inputs are supported when the selected endpoint and model accept vision messages. Capability selection is explicit; an endpoint URL alone does not establish image or schema support.

Copilot

Copilot uses direct HTTPS through IntelligenceXClient. The SDK requires no Copilot CLI, native runtime, or additional provider package. It discovers the selected model's advertised protocol and uses Responses or Chat Completions through the shared conversation engine.

using IntelligenceX.Copilot.Native;
using IntelligenceX.OpenAI;

var options = new IntelligenceXClientOptions {
    TransportKind = OpenAITransportKind.CopilotNative,
    DefaultModel = "gpt-5.4",
    CopilotOptions = new CopilotNativeOptions {
        GitHubToken = Environment.GetEnvironmentVariable("COPILOT_GITHUB_TOKEN")
    }
};
using var client = await IntelligenceXClient.ConnectAsync(options);
var models = await client.ListModelsAsync();
var turn = await client.ChatAsync("Summarize the supplied text.");
Console.WriteLine(EasyChatResult.FromTurn(turn).Text);

Choose a model returned by ListModelsAsync; availability depends on the account. Streaming, tool calls, inline images, structured-output requests, local conversation history, and usage use the same client contracts as the other native providers. Image and schema support still depend on the selected model. No tools execute unless the host supplies and runs them.

For credentials, supply GitHubToken, an asynchronous TokenProvider, or an explicit IAuthBundleStore. With no explicit source, the SDK checks COPILOT_GITHUB_TOKEN, GH_TOKEN, then GITHUB_TOKEN. An explicit store or account selection never falls back to those environment variables. AccountId is the numeric GitHub user ID represented as a string; the SDK verifies a pinned identity before inference.

For device sign-in, configure GitHubClientId for a registered GitHub app with device flow enabled, then call LoginCopilotAsync(code => ShowCode(code.UserCode, code.VerificationUri)). The host owns the UI. Supply an auth store to persist credentials; otherwise they remain in this client instance. Expiring stored credentials renew through the same registered app. Some app types also require GitHubClientSecret. GitHub app registration and account entitlement must permit Copilot access; completing sign-in alone does not establish inference entitlement.

For bounded document operations, wrap the connected client in OpenAIChatTreatmentProvider. Ephemeral = true removes local conversation state after the operation, and MaxResponseBytes bounds wire bytes including streaming frames. Native Copilot defaults to a ten-minute request deadline and a 16 MiB response limit. Hosted service retention remains a separate policy.

This transport implements the Copilot client inference protocol, which is not a versioned public GitHub REST inference API. It does not use the GitHub Copilot management API or promise protocol stability. Unsupported models and requests fail explicitly; the SDK does not install or fall back to a CLI.

Migrating from the retired Copilot clients

Replace OpenAITransportKind.CopilotCli, CopilotClient, CopilotChatClient, and the experimental CopilotDirectClient with IntelligenceXClient configured as above. Replace CopilotTreatmentProvider with OpenAIChatTreatmentProvider over that client. Remove CLI paths, launchers, installers, environment forwarding, and direct-wrapper configuration. Use copilot-native in saved configuration. Legacy copilot-cli selections are rejected rather than selecting another provider.

Existing custom IAuthBundleStore implementations remain supported for reading and saving credentials. Implement the optional IRemovableAuthBundleStore capability for persistent selective logout; otherwise logout reports NotSupportedException and the current client stops using the credential. A null account removes only the provider's accountless entry. Other providers and accounts must remain stored. The shared store accepts credentials without refresh tokens; renewal is attempted only for expiring credentials that need it.