AI agents

Connect Claude to DialSheet

Last updated September 10, 2026

You have a day of calls in DialSheet and a Claude window open. This guide connects the two, so Claude reads the transcripts and summaries, writes the recap, creates the follow-ups and drafts the emails inside your account instead of from text you pasted. It takes about three minutes in claude.ai or Claude Desktop, one command in Claude Code, and a five-line file in Cursor.

What you need

  • A DialSheet account on Pro ($29/month, then $19 per seat). The API and the MCP server are Pro features. On Free the consent page loads, explains that, and the Allow button stays off.
  • Some recorded calls. The agent reads what Call Intelligence produced for each one: the transcript, the summary, the disposition and the next step. A call nobody answered has none of those.
  • One of: claude.ai with Connectors enabled (check Settings → Connectors), Claude Desktop, Claude Code, or Cursor. Any other client that speaks MCP over HTTP with OAuth or a bearer header should work too; those four are the ones we test.

claude.ai and Claude Desktop: three clicks

  1. Open Settings, then Connectors, then Add custom connector.
  2. Name it DialSheet and paste https://www.dialsheet.io/api/mcp into the URL field. Leave the client ID and client secret fields empty. The server registers the client itself when Claude first knocks.
  3. Click Connect. A DialSheet page opens. Sign in if you are not already, read the list of what the connector gets, and click Allow. You land back in Claude with DialSheet showing as connected.

Claude Code: one command

Create a key first under Settings → API Keys (it is shown once, so copy it), then run this in a terminal. Claude Code keeps the header for you.

terminal
claude mcp add --transport http dialsheet https://www.dialsheet.io/api/mcp \
  --header "Authorization: Bearer ds_your_key"

Type /mcp inside Claude Code to confirm it lists dialsheet with its tools. If it says unauthorized, the key was pasted wrong or the plan it belongs to is no longer Pro.

Cursor: a five-line file

Put this in .cursor/mcp.json at the root of the project you work from. Cursor reads it on start and shows the tools under its MCP settings.

.cursor/mcp.json
{
  "mcpServers": {
    "dialsheet": {
      "url": "https://www.dialsheet.io/api/mcp",
      "headers": { "Authorization": "Bearer ds_your_key" }
    }
  }
}

The first three things to ask it

“Write my sales recap for today in America/Chicago time.”

Claude calls the daily_recap tool, which returns the day grouped by prospect with each call’s AI summary, next steps, disposition, your notes, open tasks and open deals, and writes the recap from it. Name your time zone or “today” is read in UTC and the evening calls fall into tomorrow.

“Which prospects said callback this week? Create a task for each for tomorrow at 9.”

One list_calls call filtered to the week and the callback disposition, then one create_task per prospect. Check the task list afterwards; it is your calendar it just filled.

“Pull my longest call today and tell me where I lost them.”

get_call returns the full transcript for one call. This is the read that turns a two-line AI summary into coaching, and it is where the agent earns its keep on a bad afternoon.

The exact recap structure, and the endpoint behind it if you would rather not use an agent at all, is in the daily recap guide.

What the agent can and cannot do

It can:

  • Read every call with transcript, AI summary, disposition and next steps, filtered by day or range in your time zone.
  • Find, create and update leads, notes, tasks, tags and deals, and move deals through the pipeline.
  • Enroll leads in sequences, and send email and SMS from the accounts you connected.
  • Import a pasted list of up to 500 leads in one go and tag them.
  • Give you the day, or the team’s day if you are the org admin, as one recap.

It cannot:

  • Place or answer calls, or listen while you are on one. It works after the call, from the recording.
  • See a teammate’s calls unless you are the organisation admin and ask for the team recap.
  • Run on the Free plan.
  • Know anything about a call that was not recorded. No recording, no transcript, no summary.
  • Push you a notification when a call ends. It reads when asked, or on a schedule you set on the Claude side.

Security, in plain terms

The agent is you. Every tool call runs with a key that belongs to your account and sees only what your login sees. Email goes out from your connected mailbox and SMS from your Twilio number, in your name, so the habit to build is reading a drafted message before you say send. The key never expires on its own; disconnecting means deleting it under Settings → API Keys, which takes effect on the next request. We do not train anything on your calls. What Claude retains of what it reads is set by the Claude plan you are on, not by DialSheet.

If something goes wrong

The consent page says upgrade

The account you signed in with is on Free. The API and MCP server are Pro ($29/month for the first seat). Upgrade in Settings → Billing, then click Connect again.

“Unauthorized” in Claude Code or Cursor

The key is wrong, or the account it belongs to is no longer Pro. Keys are checked against the plan on every request. Make a fresh key under Settings → API Keys and paste it again.

Claude Desktop has no Connectors menu

You are on a build from before custom connectors. Update it. If you cannot, the bridge below works from the old config file: npx -y mcp-remote https://www.dialsheet.io/api/mcp

The recap comes back empty

The day is computed in the zone you name and defaults to UTC. Say the time zone in the prompt (“in America/Denver time”) or pass tz to the tool, and check that the calls in question were actually recorded.

Common questions

Does DialSheet have an MCP server?

Yes. It is hosted at https://www.dialsheet.io/api/mcp and needs no install. claude.ai and Claude Desktop connect through Settings → Connectors → Add custom connector and an Allow click; Claude Code and Cursor connect with an API key from Settings → API Keys. It is part of the Pro plan ($29/month, then $19 per seat); the free tier has the dialer but not the API or the MCP server.

Do I have to install or run anything on my computer?

No. The server runs on dialsheet.io. In claude.ai and Claude Desktop you paste one URL and click Allow. Claude Code needs one command and Cursor a five-line config file, both of which are on this page. A local package exists for clients that cannot send a header to a remote server, and almost nobody needs it.

What can Claude do once it is connected?

Read every call with its transcript, AI summary, disposition and next steps, filtered by day or date range in your time zone; ask for a daily recap grouped by prospect; find, create and update leads, notes, tasks, tags and deals; enroll leads in sequences; and send email and SMS from the accounts you connected. It acts as you, in your account.

Can Claude make or answer calls?

No. It reads what happened on calls after they are recorded and transcribed, and it can create the callback task, but the dialing stays with you in the browser dialer. Calls that were never answered have no transcript and no summary, so the agent sees them as dials with a status and nothing more.

Is it safe to give an AI access to my CRM?

The connector holds an API key scoped to your own account, listed in Settings → API Keys as "Claude (MCP connector)". Delete that key and the connection is dead. Everything it sends goes out from your mailbox or your Twilio number in your name, so read a drafted email before you tell it to send. What Claude does with what it reads is governed by your Claude plan, not by us.

How much does it cost?

Nothing on top of Pro, which is $29/month, then $19 per seat on your own Twilio account. The server itself is free to call as often as you like; the tokens Claude spends reading your transcripts are on your Claude plan. There is no per-seat charge for the connector: every Pro user on a team can connect their own.

Menu names are as of September 10, 2026. The Connectors menu belongs to Anthropic and moves between releases; the server URL does not. If a step on this page no longer matches what you see, tell support which client and version you are on.