Chat Runtime Modes (ChatGPT, Copilot, LM Studio, Azure, Other Providers)

IX Chat supports three runtime transports in the desktop app:

Important:

• Runtime selection and sign-in are related but distinct.
• API keys are used only in compatible-http.
• copilot-cli uses Copilot subscription authentication, not API keys.
• You can stay signed in to ChatGPT while using non-native runtime modes.

Chat App (Runtime Selection)

For the WinUI desktop app (Build/Run-ChatApp.ps1):

1. Open Options -> Runtime.
2. Use one of the primary actions:
   • Use ChatGPT Runtime (native)
   • Use LM Studio Runtime (compatible-http + LM Studio base URL)
   • Use Copilot Subscription (copilot-cli) 3. Runtime actions apply immediately in-process and show an in-progress state; runtime switches do not auto-restart the sidecar. 4. Click Refresh Models after switching runtime or changing endpoint details when you want a forced re-probe. 5. Use Show Advanced Runtime when you need explicit transport/base URL/API key/manual model overrides.

Advanced presets:

• Use LM Studio Runtime: http://127.0.0.1:1234/v1
• Use Ollama Runtime: http://127.0.0.1:11434
• Use Copilot Endpoint (API key): https://api.githubcopilot.com/v1 (compatible-http path, API key required)

Notes:

• Refresh Models applies pending runtime field changes first, then refreshes discovery.
• For compatible-http and copilot-cli, the app auto-selects the first discovered model only when no model is set. Explicit/manual model IDs are preserved.
• If you choose Manual model input in Discovered models, type the exact model id in the Model field and click Apply Runtime.
• Leaving API key empty keeps the currently saved compatible-http key unchanged.
• Use Clear Saved API Key to remove the saved compatible-http key from the active profile.
• Compatible HTTP auth modes: Bearer (API key), Basic (username/password), or None.
• Anthropic/Gemini entries are bridge presets (experimental): IX Chat targets your configured compatible endpoint and does not do native subscription login itself.
• ChatGPT sign-in and account switching are native-runtime controls. In non-native modes (compatible-http/copilot-cli) the menu hides sign-in actions and chat is not gated on ChatGPT login state.
• The panel shows active runtime/model status so you can confirm what is currently used.
• While a runtime switch is running, runtime buttons are temporarily disabled to avoid duplicate apply requests.

Reasoning controls:

• native and most compatible-http presets expose reasoning controls (minimal|low|medium|high|xhigh) as provider hints.
• copilot-cli currently does not expose reasoning controls.
• Experimental Anthropic/Gemini bridge presets currently use provider-default reasoning.

Model Discovery and Runtime Detection

• Auto Detect probes localhost endpoints for LM Studio (127.0.0.1:1234) and Ollama (127.0.0.1:11434).
• If localhost is not available, IX Chat also probes your currently configured compatible-http base URL.
• This helps external endpoints (for example remote LM Studio or Azure/OpenAI-compatible gateways) reflect real availability.
• For LM Studio endpoints, IX Chat also reads LM Studio catalog metadata (state, quantization, architecture, context lengths, capabilities) when available and surfaces it in the runtime model picker/state note.

Troubleshooting:

• Couldn't connect to local runtime after startup: Timed out waiting for service pipe. indicates the selected runtime transport did not become reachable in time.
• For LM Studio, verify Local Server is running and http://127.0.0.1:1234/v1/models returns data before retrying.

Security Model

• http:// is rejected by default.
• For loopback http://127.0.0.1/... or http://localhost/..., you must opt in with --openai-allow-insecure-http.
• For non-loopback http:// (not recommended), you must opt in with --openai-allow-insecure-http-non-loopback.

Chat Host (Recommended For Local Use)

The simplest way to dogfood locally from this repo is the REPL host (IntelligenceX.Chat.Host).

Ollama example:

pwsh ./Build/Run-Chat.ps1 `
  -AllowRoot C:\Support\GitHub `
  -OpenAITransport compatible-http `
  -OpenAIBaseUrl http://127.0.0.1:11434 `
  -OpenAIAllowInsecureHttp `
  -Model llama3.1

LM Studio example:

pwsh ./Build/Run-Chat.ps1 `
  -AllowRoot C:\Support\GitHub `
  -OpenAITransport compatible-http `
  -OpenAIBaseUrl http://127.0.0.1:1234/v1 `
  -OpenAIAllowInsecureHttp `
  -Model <your-model>

Notes:

• -OpenAIBaseUrl can be either http://host:port or http://host:port/v1. IntelligenceX normalizes it to /v1/ internally.
• For local providers, -OpenAIApiKey is usually optional (many servers ignore it).

Copilot subscription example:

pwsh ./Build/Run-Chat.ps1 `
  -AllowRoot C:\Support\GitHub `
  -OpenAITransport copilot-cli `
  -Model gpt-5.3-codex

Chat Service (When Embedding In An App)

IntelligenceX.Chat.Service is a JSONL/named-pipe service designed to be started/managed by a UI (desktop app) or other supervisor.

Example:

dotnet run --project IntelligenceX.Chat/IntelligenceX.Chat.Service/IntelligenceX.Chat.Service.csproj --framework net10.0-windows -- `
  --openai-transport compatible-http `
  --openai-base-url http://127.0.0.1:11434 `
  --openai-allow-insecure-http `
  --model llama3.1 `
  --allow-root C:\Support\GitHub

Copilot subscription service example:

dotnet run --project IntelligenceX.Chat/IntelligenceX.Chat.Service/IntelligenceX.Chat.Service.csproj --framework net10.0-windows -- `
  --openai-transport copilot-cli `
  --model gpt-5.3-codex `
  --allow-root C:\Support\GitHub

Tool Calling Support

IntelligenceX will send OpenAI tool definitions (tools) and expects tool calls in message.tool_calls.

Provider compatibility varies:

• Some local servers fully support tool calling.
• Others ignore tools or return tool calls in non-standard shapes.

When tool calling is not supported by the provider, the chat still works but tools will not be invoked.

Current Limitations

• Image inputs are not supported for compatible-http yet (text + tools only).