LLM and speech recognition
Screph supports local and cloud connections. Each feature uses an explicitly selected profile and does not switch to another service without changing settings.
LLM profiles
A profile contains the service type, address, model and request parameters. Supported services include OpenAI, Anthropic, Gemini, OpenRouter, OpenAI-compatible services, Ollama, LM Studio and LiteLLM Proxy.
Assign connection profiles to Screph Code, the Agent, computer vision, semantic processing, and connection checks. Choose a model suited to the task.
ChatGPT, Codex, and OpenHands connections
ChatGPT/Codex uses a separate account connection and an installed Codex. OpenCodex connects Codex to a selected compatible service and requires separate configuration. OpenHands uses a supported direct or proxy connection with a model and key. OpenHands does not use ChatGPT account sign-in.
The agent engine, role profile, and model are selected separately. Changing a connection does not transfer an existing conversation to another account. If the selected model, engine, or profile is unavailable, the reason is shown; there is no automatic substitution with a service that has different data-sharing terms.
Local and cloud processing
| Option | Data transfer |
|---|---|
| Ollama / LM Studio | The context is processed by a local service on the user's computer or on the specified local network. |
| Cloud LLM profile | Allowed text, images and context are sent to the selected provider. |
| User-configured compatible service | Data is sent to the address specified in the profile. You are responsible for deciding whether to trust that address and accept its terms. |
Keys and connection check
API keys are stored in Windows Credential Manager. Regular settings keep the service address, model, and request parameters. If the system credential store is unavailable, the key must not be stored in plain text.
Saving or checking a profile can contact the selected service to check access and get a list of models. This check does not submit the project, but the service receives normal network request data and profile credentials.
Speech Recognition
| Option | Requirements and data |
|---|---|
| Vosk | A local model for the selected language. Audio stays on your computer. |
| Yandex | A direct connection sends audio to Yandex using your configured Yandex credentials. It does not use Screph server minutes. |
| Google Cloud Speech | Audio is transmitted to Google with configured credentials. |
| OpenAI-compatible | The audio is sent to the specified compatible service. |
| ChatGPT / Codex | Experimental Live dictation through a ChatGPT account and an installed Codex. Signing in to the account alone does not confirm voice feature availability. |
Batch, Live, and recording controls
Batch recognises accumulated audio after recording ends. Live displays interim text during recording; the final text is refined when recording ends. Vosk supports both modes locally. Cloud mode availability depends on the selected service, model, and installed components.
Pause stops new audio input and allows the same session to resume; it does not release the microphone. Finishing a recording waits for the final text. Cancel stops the current request but cannot recall audio already sent to an external service.
Recording requires a selected, available microphone. The service handles one microphone session at a time; another request does not start recording while the device is busy. Provider, model, or mode changes apply to the next session, preserving the settings of the current recording.
Voice annotation saves text on the selected object, while an Agent voice request goes to the chat. Additional semantic processing of annotations depends on its settings. Recognized text can be reviewed and corrected; ChatGPT dictation and cloud modes remain experimental.
Settings
For speech recognition, the provider, microphone, language, and mode are selected in Voice settings. Assigning an LLM role does not configure the microphone or replace the speech recognition model.
- Create a profile and specify your service, address and model.
- Save the key via system storage if required.
- Check your connection and assign the profile to the correct role.
- Test the connection with non-sensitive data before using it on a working project.
If the connection doesn't work
- Check the address, model, key, network access and provider account restrictions.
- For a local service, make sure it is running and the model is loaded.
- For Vosk check the local model language and path.