One address to connect
Streamable HTTP with OAuth: no key to paste, no server to host. Works with Grok Bot, Claude, ChatGPT, and Cursor today.
An agent can browse, write, and call an API. It cannot ring the tyre shop that has no API, the clinic that only answers a phone, or the landlord who never replies to email. CosVoice is a phone number and an email address your agent uses over MCP, with the call handled for it and the outcome handed back as fields it can read.
Wire up a telephony provider, rent a number, run a media server, glue a speech model to it, write the turn-taking, handle interruptions, and keep the whole thing alive. Then build the part that decides what to say.
Paste https://cosvoice.com/mcp into your agent and press Allow. The number, the voice, the conversation, and the write-up already exist. Your agent supplies the intent in a sentence.
Streamable HTTP with OAuth: no key to paste, no server to host. Works with Grok Bot, Claude, ChatGPT, and Cursor today.
The same tools over HTTPS with a bearer key, described in a public OpenAPI document at /openapi.json that an agent can read without signing in.
place_call dials out from a plain-English brief. The same number answers when someone rings back, with your standing instructions in place.
get_call returns the outcome, the details the line captured, action items, and the full transcript. Nothing to transcribe yourself.
Rather than polling, give us a webhook and we start your agent when a call finishes, mail lands, or the owner texts an instruction.
The line says it is an assistant when asked, never reads a card number, and will not agree to a charge. You do not write those prompts.
Hi, this is Lloyd, calling on behalf of Graham. I'm calling about a quote for a tankless water heater, installed.
Sure. Four thousand eight hundred fifty, including the permit and haul-away.
Thank you. How long does that price hold, and how soon could you start?
Thirty days, and we could start a week from Monday.
Your agent asked once: place_call with a purpose and what to accept instead. It gets back an outcome, the amount, the validity, the start date, and the transcript, without handling a single audio frame.
Recorded calls, both sides, with every word written out and the note the owner gets.
Listen and read along →
Listen and read along →
Listen and read along →
Starter is $29 a month with 60 minutes of calls. Plus is $69 with 200, Pro $149 with 600, and 25 cents a minute beyond that. A 40-second call costs about 20 cents of your minutes, and there is no per-call fee.
A remote MCP server at https://cosvoice.com/mcp, over Streamable HTTP, authorised with OAuth 2.1 and PKCE. There is no key to paste and the owner can revoke it in Settings.
Use the REST API with a bearer key. The machine-readable description is at cosvoice.com/openapi.json, readable without signing in, so an agent that writes its own connector can start there.
Yes. Connect whichever assistant you are using; they all reach the same number, the same history, and the same standing instructions.
No. The number, the carrier side, and the voice are included. Nothing to host and nothing to deploy.
Read next: The tools, the REST API, and the webhooks · The page written for your agent · Agents that build their own connector
Pick a number, save the contact, make the first call. About a minute.