First voice MCP provider — Russia-hosted, FZ-152 compliant

Voice channel for any AI agent over MCP

Plug telephony into Claude Desktop, AGONTS, LangChain or your own agent in 5 minutes. One API key, open MCP standard, JSON-RPC at /mcp/v1/jsonrpc. Tools for calls, SMS, callbacks, outbound campaigns — straight into your LLM's context. Servers in Russia, FZ-152 ready.

Get started — 5 minutes View JSON-RPC
No card to sign upSelf-serve activationMCP-2026 compliant

What it looks like inside Claude Desktop

A real conversation: the user asks Claude to call a customer about tomorrow's appointment. Claude picks the make_call tool from the Stexa MCP server, sees the tool result, and reports back to the human.

Frame from a real Claude Desktop + Stexa Voice MCP session. Want a live demo? Sign up — sandbox DID is yours in a minute.

Compatible with any AI agent

Stexa Voice MCP is the JSON-RPC server side. The client can be any tool that speaks MCP. Three most common paths for developers and integrators.

Claude Desktop in 1 minute

Add Stexa to claude_desktop_config.json and the LLM gets the make_call tool right in its context. No backend, no webhook, no hosting on your side.

{
  "mcpServers": {
    "stexa-voice": {
      "command": "npx",
      "args": ["-y", "@stexa-ai/mcp-client"],
      "env": {
        "STEXA_API_KEY": "sk_live_..."
      }
    }
  }
}

AGONTS workspace plugin

Drop @stexa-ai/agonts-plugin (npm) into your tools array and your AGONTS agent learns how to call, SMS, and schedule.

// agonts.workspace.ts
import { StexaVoiceMCP } from "@stexa-ai/agonts-plugin";

export const agent = defineAgent({
  name: "Sales bot",
  tools: [
    StexaVoiceMCP({
      apiKey: process.env.STEXA_API_KEY,
    }),
  ],
});

LangChain / custom agent

One-line connection via langchain-mcp's MCPClientToolkit. Works with any LLM — Claude, GPT, YandexGPT, GigaChat, local.

from langchain_mcp import MCPClientToolkit

toolkit = MCPClientToolkit(
    server_url="https://stexa.ru/mcp/v1/jsonrpc",
    bearer_token=os.environ["STEXA_API_KEY"],
)

agent = create_react_agent(llm, toolkit.get_tools())

Claude Desktop, AGONTS, Cursor are trademarks of their respective owners. Stexa AI is not affiliated and works through the open Model Context Protocol standard.

How it works

Stexa Voice MCP is one JSON-RPC endpoint that implements the MCP-2026 spec. No webhook servers on your side, no queues. Three steps from sign-up to your AI agent's first live call.

Get your API key

Email sign-up, no credit card. On any voice plan the dashboard opens the “MCP Provider” section — create a key there. One key = one tenant, rotate any time.

Drop the URL into your agent

Endpoint https://stexa.ru/mcp/v1/jsonrpc plus Bearer token — into Claude Desktop / AGONTS / LangChain config. The LLM discovers tools automatically via tools/list.

The agent calls

LLM decides to invoke make_call — Stexa initiates the call via Asterisk ARI, runs the TTS script, records the conversation, returns call_id and status.

Full request example

Any JSON-aware HTTP client can talk to Stexa Voice MCP. Below — the anatomy of one make_call tool invocation with the full parameter set.

POST https://stexa.ru/mcp/v1/jsonrpc
Authorization: Bearer sk_live_a1b2c3d4...
Content-Type: application/json

{
  "jsonrpc": "2.0",
  "id": 42,
  "method": "tools/call",
  "params": {
    "name": "make_call",
    "arguments": {
      "to": "+14155550123",
      "script": "Hi! Confirming your 2 PM appointment.",
      "voice": "alyona",
      "max_duration_sec": 60,
      "record": true,
      "callback_url": "https://example.com/webhooks/stexa"
    }
  }
}

// ↩ Response 200 OK (412 ms)
{
  "jsonrpc": "2.0",
  "id": 42,
  "result": {
    "content": [{
      "type": "text",
      "text": "{\"call_id\":\"call_01HXQ7P\",\"status\":\"dialing\",\"eta_seconds\":3}"
    }]
  }
}

Full tools reference (make_call, send_sms, schedule_callback, voice_outbound), error codes, and SDKs — on GitHub.

Tech specs

No vendor lock-in: open MCP standard, documented JSON-RPC, official catalog listings (Smithery + mcp.so). Transparent roadmap for OAuth and tool expansion.

Tools / methods

  • make_callprod
  • send_smsprod
  • schedule_callbackprod
  • voice_outboundprod
  • conference_bridgeQ2 2026
  • call_transferQ2 2026

Protocol & transport

  • MCP-2026-03-26current
  • JSON-RPC 2.0 over HTTPSprod
  • TLS 1.3, HTTP/2prod
  • Bearer tokens per tenantprod
  • OAuth 2.1 PKCEprod
  • RFC 7591 dynamic client registrationprod

Security & compliance

  • FZ-152 (RU PDP law)prod
  • Data processing agreementprod
  • Russia-hosted (Yandex Cloud)prod
  • Voice recording consentprod
  • Tool-invocation audit logprod
  • FSTEC class 4 protectionprod

SDKs & clients

  • @stexa-ai/mcp-client (npm)prod
  • n8n-nodes-stexa-ai (npm)prod
  • Claude Desktop compatibleprod
  • AGONTS / Cursor compatibleprod
  • LangChain MCPClientToolkitprod
  • Python SDK (PyPI)Q2 2026
Our place in the MCP ecosystem
9,400+
MCP servers in Smithery + mcp.so registries as of May 2026
97M
MCP client installs across the ecosystem (Claude Desktop, AGONTS, Cursor)
#1
voice MCP provider for Russian-language AI agents in ecosystem catalogs (May 2026)
Smithery Listed on Smithery mcp.so Listed on mcp.so
Smithery registry mcp.so directory Claude Desktop AGONTS workspace

Sources: smithery.ai (registry statistics, May 2026), modelcontextprotocol.io. Ecosystem logos are trademarks of their respective owners, shown for informational purposes.

Часто задаваемые вопросы

What is a voice MCP provider and why does my AI agent need one?
MCP (Model Context Protocol) is an open standard for connecting AI agents to external tools — databases, search, file systems. A voice MCP provider adds telephony to that toolset: your agent can place a call, send an SMS, or schedule a callback through a single tool invocation. Stexa is the first voice MCP provider with native Russian-language support and FZ-152 compliance (as of May 2026). Out of 9,400+ MCP servers in Smithery + mcp.so registries, none target Russian-language telephony at this depth.
How much does it cost? Is there a free tier?
Sign-up is free and needs no credit card — you can configure the agent and connect your own keys right away. You can try it before paying: 7 days free with 10 calls of up to 2 minutes each — no card, on your own Yandex Cloud key and phone number. Everything except outbound campaigns. An MCP key and calls placed through the MCP tool are included — they count toward the same 10 calls. You start the trial with one click in the dashboard once your key and number are connected. After the trial, the API key and MCP tools come with any voice plan: Starter, Business, or Pro — pay for the plan in the dashboard and access turns on automatically, no sales call, no document exchange. Current prices and minute limits are shown in the dashboard.
What's the latency from tool call to the customer hearing the dial tone?
400-600 ms from JSON-RPC request received to actual dial on Asterisk — significantly lower than the 2-4 seconds typical of REST voice APIs. We achieve this through a direct MCP → Asterisk ARI bridge with no intermediate REST queue. If your LLM is mid-conversation, the agent can narrate ("hold on, dialing now…") while the call initiates — the customer never hears silence.
How does the OAuth flow work for new clients?
We implement OAuth 2.1 with PKCE plus RFC 7591 Dynamic Client Registration. Your AI agent (Claude Desktop, Cursor, n8n, custom) POSTs to /mcp/v1/oauth/register with a client_name and redirect_uri — gets back a client_id and client_secret. Then a standard PKCE-protected authorization code flow exchanges the user consent for an access token. Discovery via /.well-known/oauth-authorization-server (RFC 8414) and /.well-known/oauth-protected-resource (RFC 9728).
What about data security and compliance?
Per-tenant Bearer tokens (rotatable in dashboard) plus OAuth 2.1 scopes for granular access. All traffic over TLS 1.3. Servers are hosted in Russia (Yandex Cloud Moscow). For PDP compliance under Russian federal law FZ-152, a data processing agreement is signed on Business tier and above. Every outbound call announces the recording consent and captures the customer's response — stored 3 years as a legal artifact. We do not store call audio outside Russia under any circumstance.
Which LLM providers are supported?
Stexa Voice MCP is the server side of the transport. The client can be any MCP-aware LLM: Claude (Anthropic), GPT-4/5 (OpenAI), YandexGPT, GigaChat, local models via LM Studio. Stexa internally uses YandexGPT Pro to understand the customer's speech during the call (98% accuracy on Russian), but when you use the MCP channel, LLM orchestration remains on your side — we only execute the tool invocations: make_call, send_sms, etc.
How is this different from Twilio or Vonage REST APIs?
Twilio / Vonage follow a REST + webhook model: you write a backend, register a webhook URL, Twilio POSTs JSON on each event, you respond with TwiML. That requires backend infrastructure and does not integrate directly into an LLM agent. Stexa's MCP model: your AI agent itself is the client, no backend between it and telephony. In Claude Desktop you add one line to config.json and the LLM gets the make_call tool in its context. No webhook server. For AI agents this is an on-ramp without code.
Get started in 5 minutes

Connect telephony to your AI agent today

Sign up without a card, activate without a sales call. The API key and MCP tools come with any voice plan. Full docs, sandbox DID, and Telegram support included.

Start in 5 minutes View on GitHub