Profiles, models, and activation

A profile groups the assistant’s identity, selected intelligence provider and model, voice behavior, and activation mode. Use profiles when the same agent computer needs distinct working styles.

Create and select profiles

Open Profiles Studio and select NEW. Give the profile a specific name and select it from Preset when you want it active. The default profile cannot be deleted, and at least one profile must remain.

Wait for the synchronization indicator before closing the studio after important changes.

Write the system prompt

The System Prompt is durable guidance injected into conversations for that profile. Prefer short, testable rules:

• role and areas of responsibility;
• response style and level of detail;
• important sources or tools to prefer;
• actions that must be confirmed;
• facts the agent must not assume.

Do not use the system prompt as secret storage or as a substitute for tool permissions.

Choose provider and model

Select an available Provider, then choose a Model from the provider-specific searchable list. The provider must have a valid managed or BYOK credential. A model can disappear when the provider catalog changes or an account lacks access.

Use a fast model for routine conversation and tool routing. Use a more capable model for difficult analysis or planning when the additional latency and cost are justified.

Configure voice and speech

Turn Speech Output on to generate spoken responses. Native audio providers select a provider voice and ignore local TTS settings. Non-native providers select a TTS Engine and Voice.

Voice Studio profiles appear only when the service is connected and responsive. See Voice Studio to create and maintain them.

Choose activation

Push to Talk requires an explicit microphone action. Wake Mode listens for configured wake names and retains a follow-up window. Start with push-to-talk in shared spaces, then enable wake mode after confirming microphone and echo behavior.

Test a profile

After changing a provider, model, or voice:

1. send a short typed request;
2. ask for one tool-free response;
3. test speech input and output;
4. run a harmless read-only tool;
5. confirm the active profile label before starting important work.