Understand models, routing, and context
Vox currently defaults to xAI (Grok). Its provider and model are deployment configuration, not hard-coded product assumptions. Wirevox retains an OpenAI-compatible path, so an operator can switch Vox back to OpenAI later without replacing the capability and approval architecture.
The configured model writes and reasons about the response. It does not authorize itself. Wirevox independently selects the available capabilities, validates arguments, checks tenant permissions, and controls every mutation.
Request modes
Vox classifies each turn into the smallest useful mode:
| Mode | Typical request | Live tools |
|---|---|---|
| Casual | “Hi” or “Thanks” | None |
| Capability overview | “What can you do?” | None; category-level answer |
| Guide | “How do I write a system prompt?” | Normally none |
| Workspace read | “How many credits remain?” | Relevant authorized reads only |
| Action | “Create a voice agent” | Reads needed to validate, then approval-backed writes |
| Research | “Look up this company” | Public web and selected-page tools |
This routing is why Vox should not search the web for ordinary greetings or unrelated follow-ups. A short continuation such as yes, proceed, or continue inherits the recent actionable request so Vox can finish the correct workflow.
Context is bounded
Vox does not send the entire workspace, complete thread, or full capability catalog to the model on every turn. The selected mode controls recent messages, context size, visible capabilities, tool rounds, output length, and the credit reservation.
The largest action and research modes currently include up to 20 recent messages and 24,000 context characters. Smaller conversational modes use substantially less. Long tool results and attachment analyses are truncated before they enter model context.
Page and research context
When appropriate, Vox can receive bounded context about the current Wirevox page. Page context helps interpret references such as “this agent,” but it is not an authorization source and does not replace an exact live read.
Web research artifacts are stored with the conversation and a bounded portion can be reused in later turns. This avoids repeating the same search unnecessarily. Research context is not permanent workspace memory and is not a knowledge base unless you approve an ingestion action.
Provider-neutral design
Vox capabilities are defined independently from the model provider, with schemas, effects, permissions, versions, and approval requirements. Wirevox exposes this policy-controlled capability layer through an authenticated MCP server for the Wirevox ChatGPT connector. The connector applies the signed-in user's workspace permissions and requires Wirevox approval cards for supported changes.