Sarvam brings voice-agent deployment into coding assistants

Sarvam has released the Voice Agents MCP, which connects Cursor, Claude Code, Windsurf, and other compatible clients to its managed voice platform. Developers can ask a coding assistant to configure an agent, assign a phone number, launch an outbound campaign, and retrieve call results without manually wiring each API call.

Model Context Protocol, or MCP, gives AI clients a standard way to discover and invoke external tools. Sarvam’s server turns voice-agent operations into those tools, allowing the coding assistant to translate a plain-language request into a sequence of platform actions.

One protocol, two layers

Sarvam previously released a general MCP server for its public APIs, including speech recognition, text-to-speech, translation, language models, and document vision. That server exposes individual inference capabilities to clients such as Claude Desktop, Cursor, Windsurf, and Zed.

The Voice Agents MCP operates at the workflow level. It manages the lifecycle of a deployed phone agent through higher-level tools:

From instruction to campaign

Sarvam Voice Agents, formerly called Samvaad, is now available as a self-service product. According to the company, the platform was developed through enterprise deployments involving noisy phone lines, interruptions, code-mixed speech, mid-sentence language changes, and large concurrent call volumes.

A compatible coding assistant can use the new MCP tools to complete the following sequence:

1. Create an agent with a prompt, language, voice, and external tools.
2. Attach an existing phone number or rent one through Sarvam.
3. Upload a contact cohort and configure an outbound campaign.
4. Start calls after review and approval.
5. Return call logs, transcripts, outcomes, and analytics to the client.

These operations map to the platform’s existing building, testing, deployment, campaign, and reporting features. MCP supplies a conversational control surface for them inside the development environment.

Managed orchestration changes the trade-off

A custom voice stack commonly combines speech recognition, a language model, speech synthesis, telephony, campaign scheduling, and observability. Each integration adds authentication, error handling, latency, retries, and version maintenance. Sarvam manages that orchestration and exposes the assembled workflow through agent-level tools.

The approach reduces integration work while increasing dependence on Sarvam’s platform, pricing, supported languages, and operational controls. Teams still own agent instructions, business integrations, contact data, testing, compliance, and approval policies.

Sarvam lists Voice Agents in 11 Indian languages, with pricing starting at ₹3 per minute. An API key from the Sarvam dashboard is required to connect the MCP server to an account.

Best fit: Indian-language calls

The clearest use cases are sales, collections, customer support, and bookings involving Indian languages or code-mixed speech. Sarvam’s Saaras speech-recognition and Bulbul speech-synthesis models target those conditions, while the managed platform adds memory, business tools, phone provisioning, testing, and analytics.

Teams needing full control over telephony vendors, model routing, data residency, or low-level orchestration should evaluate whether the managed abstraction exposes the required controls. The MCP interface simplifies supported workflows but does not expand the underlying platform’s capabilities.

Claims and constraints to test

• Latency: Sarvam advertises sub-500 millisecond latency. Measure the complete call path with the intended carrier, region, tools, and agent configuration.
• Scale: The cited figure of more than 350 million conversations comes from Sarvam and lacks independent audit.
• Tool execution: Reliability depends partly on the coding assistant selecting and sequencing the correct MCP tools. Multi-step campaigns need prompt testing and explicit approval gates.
• Retries: Verify idempotency and retry behavior before allowing an assistant to create contacts or start paid campaigns.
• Data handling: Contact files, transcripts, and call recordings may contain personal information. Review retention, access controls, consent, recording disclosures, and applicable telecom rules before production use.

Connect it without guesswork

Developers need a Sarvam account, an API key, and an MCP-compatible client. The safe setup sequence is:

1. Create an API key in the Sarvam dashboard.
2. Use the Voice Agents configuration published on the launch page.
3. Store the key in an environment variable or secret manager instead of committing it to a repository.
4. Confirm which tools the client can invoke and require approval for actions that place calls or incur charges.
5. Run a small test cohort, inspect transcripts and outcomes, then increase volume.

The package name sarvam-mcp appears in examples for Sarvam’s general API server. Developers should use the Voice Agents MCP configuration for agent lifecycle tools instead of assuming the general package exposes the same operations.

A practical test prompt could ask the assistant to create a Hindi collections agent from leads.csv, display the proposed configuration and cohort size, wait for confirmation, place the calls, and summarize payment commitments from the resulting transcripts. That workflow captures the release’s main value: coordinated voice-agent operations through the coding client developers already use.