Give Claude a real phone line with the Wixzel Voice MCP server

What do you need before you start?

• A Wixzel Voice account with credit. The minimum top-up is $5, and there is no free trial credit.
• A SIP trunk at a carrier you already use (Twilio, Telnyx, Plivo, Vonage, Bandwidth, Exotel, Vobiz or any SIP provider) with a phone number on it. Wixzel Voice does not sell or port numbers.
• Claude Code, or claude.ai or Claude Desktop with custom connectors available on your plan.

How do you build this on Wixzel Voice, step by step?

1. 01Add the server to Claude Code

   The hosted server lives at https://mcp.voice.wixzel.com/mcp and signs you in with OAuth, so no API key is pasted anywhere. Add it from a terminal:

   shellCopy

   claude mcp add --transport http wixzel-voice https://mcp.voice.wixzel.com/mcp

   Then type /mcp inside Claude Code, choose the wixzel-voice server and pick Authenticate. Your browser opens the console, you sign in, untick any scope you do not want the client to have, and approve. Claude Code receives an API key of its own, limited to the scopes you approved.

2. 02Or add it to claude.ai or Claude Desktop

   In Settings, open Connectors and add a custom connector with the same URL, https://mcp.voice.wixzel.com/mcp. Press Connect, sign in to the console and approve the scopes. The connector gets its own key, and you can disconnect it from the AI clients page in the console at any time; the key stops working at once.

   Clients that launch servers locally from an mcpServers JSON config, such as Cursor or Windsurf, run the same server from npm with an API key in the environment:

   shellCopy

   claude mcp add wixzel-voice -e WIXZEL_API_KEY=wv_live_... -- npx -y wixzel-voice-mcp

   jsonCopy

   {
     "mcpServers": {
       "wixzel-voice": {
         "command": "npx",
         "args": ["-y", "wixzel-voice-mcp"],
         "env": { "WIXZEL_API_KEY": "wv_live_..." }
       }
     }
   }

3. 03Let the quickstart prompt set up your account

   The server ships a quickstart prompt, which Claude Code shows as a slash command. It takes an optional goal, such as "confirm dental appointments":

   Claude CodeCopy

   /mcp__wixzel-voice__quickstart

   Claude first lists what already exists (list_sip_trunks, list_phone_numbers, list_agents, get_balance) so nothing is duplicated. Then it works in the order the server's operating guide sets: list_engines to see what is available and what it costs, create_sip_trunk with the carrier credentials you give it, create_phone_number on that trunk, and create_agent with a short spoken-style prompt, an opening line and a caller ID. It asks you for credentials rather than inventing them, and it stops before anything that dials.

   Two settings only you can change live at your carrier. Allowlist the platform_ip the new trunk returns, so your carrier accepts outbound calls from it, and point the carrier's inbound route at the trunk's origination_uri. Claude reads both values back to you, and check_sip_trunk_status shows whether the trunk is reachable.

4. 04Ask for a call, and approve it

   With an agent and a number in place, ask in plain words:

   promptCopy

   Call +14155551234 with the Support agent. Ask whether tomorrow's 10:00
   delivery slot still works, and tell me what they said.

   Claude picks place_call. That tool is marked as one that spends money and rings a real phone, and its description tells the model to confirm with you first, so the client asks for your approval before it runs. Each place_call carries an idempotency key; a retry is safe only when Claude passes the same idempotency_key again, which the tool description asks it to do.

   The voice agent you created holds the conversation, not Claude. Claude then polls get_call until the call ends and reads you the transcript, the status and cost_micros, the cost in millionths of a dollar.

5. 05Know which tools ask before they run

   Wixzel Voice marks four tools as spending money or reaching the real world: place_call dials a number, test_call_agent rings a number and speaks one phrase, start_campaign dials every lead on a campaign, and create_topup starts a payment. Tools that only read, such as get_call or list_usage_events, are marked read-only, so a client can run them without asking.

   create_topup never pays by itself: it returns a checkout link that a person has to open and complete.

   Scopes are the second fence. A client that asks for nothing specific is offered the standard set for agents; billing:write, which starts a top-up, and api_keys:write, which mints keys, are never offered by default and have to be granted deliberately. There is no admin scope.

6. 06When a call fails, ask why

   Two more prompts ship with the server:

   Claude CodeCopy

   /mcp__wixzel-voice__diagnose_call
   /mcp__wixzel-voice__spend_report

   diagnose_call takes a call id, reads the record's failure_code and failure_reason, its usage events, the trunk status and recent SIP logs, and tells you whether the fix is at your carrier or in the agent. It places no calls while it works. spend_report turns a period's usage into dollars by component and by call.

What does this build cost on Wixzel Voice?

Which Wixzel Voice limits apply to this build?

• Phone calls need your own SIP trunk and a number on it. Wixzel Voice does not sell, port or hold numbers.
• The approval before a paid tool is shown by the client. The server marks those tools and tells the model to confirm; a client set to approve tools automatically will not ask, so leave automatic approval off for this server.
• Five calls can be live at once per account by default. The limit is raised on request.
• A test call is a real call: it rings a real phone over your carrier and the meter runs. There is no free test mode.
• Wixzel Voice is in alpha. Behavioural changes ship behind a dated Wixzel-Version header.