Dev.to AI πŸ€– Ai πŸ‘ 0 πŸ“– 2 min read

I asked Claude to get my soundbar fixed. It took six phone calls.

My Sony soundbar's subwoofer stopped pairing. Out of warranty, one failed repair already. The fix is a technician visit, and getting one means phone calls: the brand helpline, a service centre, another service centre, a

My Sony soundbar's subwoofer stopped pairing. Out of warranty, one failed repair already. The fix is a technician visit, and getting one means phone calls: the brand helpline, a service centre, another service centre, a technician.

So I gave Claude a phone. Dial is a remote MCP server: add https://dial.behaf.co/mcp to Claude, Claude Code or Cursor, and your assistant can call Indian businesses, sit through menus and hold, talk to whoever answers, and bring back the outcome and a transcript.

Here is the unedited run, straight from list_calls.

Call 1, the brand helpline. Never placed. prepare_call came back first: "Which identifier do they know you by? (registered mobile, order ID…)" plus seven questions the helpline was likely to ask (serial number, invoice, previous job sheet…). That readiness check is free and is the part I didn't expect to like: it front-loads the questions you'd otherwise fumble on the line.

Call 2, a Sony service centre. Answered, then dropped right after the greeting. Status failed, 14 s of talk.

Call 3, same centre, second try. 51 s. They don't do home visits from that branch and gave the number of another centre. Outcome partly, with the referral in the result.

Calls 4 and 5, the referred centre. Busy, twice. no_answer, nothing billed.

A refused call. I tried a punchier opening, with the AI disclosure squeezed into one short phrase. Dial refused before dialling: "The request impersonates … a real person." Strict, maybe too strict, but it errs the right way: Dia always says she's an AI assistant calling for you.

Call 6, a repair service. 141 s. Result done: home visit Saturday 2 PM, visit fee Rs 200, booking reference 901, name on booking set. The structured outcome came back as JSON (booked: true, time: "14:00", fee_inr: 200, reference: "901"), which is what makes this useful to build on.

And one more thing I liked. On a customer-care call the business read out an email address and asked us to send documents. Dial didn't act on it: the outcome came back as needs_user_decision with the flag untrusted_instruction_removed. Call content is treated as untrusted data, never as instructions, so a voice on the line can't steer your agent.

What this run says

  • Phone errands in India are messy: drops, busy lines, referrals. The agent's job is to keep going and report honestly (failed, partly, done), not to pretend.
  • People hang up on AI calls. Two of our connected calls ended fast after the disclosure. We're testing shorter openings; disclosure stays.
  • The useful bit for developers is the structure: readiness check β†’ call β†’ poll β†’ typed outcome β†’ transcript. You can chain calls (the referral from call 3 fed calls 4 and 5).

Try it

  • Claude: Settings β†’ Connectors β†’ Add custom connector β†’ https://dial.behaf.co/mcp
  • Claude Code: claude mcp add --transport http dial https://dial.behaf.co/mcp
  • Cursor: { "mcpServers": { "dial": { "url": "https://dial.behaf.co/mcp" } } }
  • Setup page + 42 s video: https://www.behaf.co/mcp Β· Docs: https://www.behaf.co/dial-docs

Rs 5 a minute + GST, billed by the second, from your own prepaid wallet. Indian (+91) numbers, person-to-business errands only. Tell me what breaks.

πŸ“° Read the original article on Dev.to AI

Originally published by Dev.to AI. Aggregated on AIWithGhost for educational purposes β€” full credit and traffic to the original publisher.