Sarvam gives voice agents a workspace catalogue of tools with per-agent controls

Sarvam added a workspace tool catalogue for voice agents with per-agent controls.

**Dek:** The October 31 release lets a workspace authorize an MCP integration, then choose the actions each voice agent can call and how each argument is filled.

Sarvam added MCP integrations to Voice Agents on October 31. A workspace can browse the integration catalogue and authorize a listed tool through the vendor’s consent screen. It can also enter a custom MCP server endpoint. In the agent editor, the builder selects the actions the individual agent is allowed to invoke. For each parameter, the builder can supply a fixed value, use a stored agent variable or let the conversation provide it. [Sarvam October 31 changelog](https://docs.sarvam.ai/changelog/2026/10/31)

Connecting a tool is one step. Assigning its actions is another.

The workspace connection makes a tool available to the workspace. The agent configuration decides which actions that particular agent may call. The split is useful for teams with more than one voice workflow. A public information agent might be allowed to look up a service location. A scheduling agent could have an additional action for booking a visit. Both can use the same integration without exposing the same action list to every caller. The example illustrates the documented configuration model. It does not establish that a particular scheduling vendor is in Sarvam’s catalogue. [Sarvam October 31 changelog](https://docs.sarvam.ai/changelog/2026/10/31)

The choice of parameter source is equally important. A fixed value can identify a queue or tenant that a public caller should never be able to change. A variable can carry a value the application verified earlier. A value derived from speech is useful for a preferred appointment date, but the same approach would be fragile for an unverified account identifier. Digits can be misheard, names can sound alike and a caller may change their mind halfway through a sentence. The design decision belongs at the point where speech becomes an external action.

What makes a voice integration safe to use

A sensible acceptance check would start with one agent and one read-only action. Try an ordinary request, an incomplete request and a spoken instruction that attempts to make the agent choose a different action. The service should enforce its own authorization and validate the actual arguments. For an action that changes a booking or sends information, the voice flow can read back the important details and ask for confirmation before invoking the tool. That creates a chance to correct an error while the caller still has context.

A custom MCP endpoint deserves the same review as any other workspace integration. Operators need to know who controls the server, what information it receives and what happens when it times out. They can choose a separate text confirmation when a spoken prompt is hard to hear or repeat. Those are deployment considerations, separate from Sarvam’s feature announcement. The release provides a way to configure the connection and action list. Its value in a production workflow depends on how the connected service and the agent enforce those choices.