A power dialer with an MCP server
Every dialer now advertises an AI feature, and most of it is a summary you read inside their app. An MCP server points the other way: your own assistant reaches into the dialer, reads the calls and does the work. This page is what that means for a cold-calling team, what ours exposes tool by tool, how it is built, and where it is the wrong tool.
What an MCP server is, for a sales team
MCP, the Model Context Protocol, is an open standard Anthropic published in November 2024. A client such as Claude connects to a server, asks it what tools it has, and calls them as the conversation needs. The practical difference from every AI feature you have seen inside a dialer is direction. Their summary arrives in their interface, and you carry it out by hand. With an MCP server, the assistant you already pay for reads the calls itself, decides which prospect needs a task, and creates it. The dialer becomes a set of hands for the model rather than another screen for you.
For a cold-calling team the interesting tools are the ones about calls, because that is where the only information nobody else has lives: what the prospect actually said. A CRM export tells an agent that a call happened. A transcript tells it why the prospect said no.
What ours exposes
| Area | Tools | What the agent gets back |
|---|---|---|
| Calls | list_calls, get_call, daily_recap | Every call with transcript, ai_summary, ai_disposition, ai_next_steps, coaching notes and your own outcome; filters for day, range, time zone and “has a transcript”; a recap of the day grouped by prospect with open tasks and deals attached. |
| Leads | search_leads, get_lead, create_lead, bulk_create_leads, update_lead, delete_lead, tag_lead, untag_lead | Search by name, company, email, phone or status; import up to 500 at once; tags created on the fly. |
| Follow-ups | list_notes, create_note, list_tasks, create_task, update_task, complete_task | Notes with a call outcome; tasks with a due date, type and priority, tied to a lead. |
| Pipeline | list_opportunities, create_opportunity, update_opportunity_stage | Deals by stage with value and probability. |
| Outreach | list_sequences, enroll_lead_in_sequence, send_sms, list_sms, list_email_accounts, send_email, list_emails | Enrollment into your sequences; email from a connected mailbox; SMS from your Twilio number; the history of both. |
| Overview | list_tags, create_tag, get_stats | Leads by status, pipeline by stage, task completion, call and email counts. |
30 tools as of September 10, 2026, plus a daily_sales_recap prompt that walks the agent through the recap. Every tool maps to a REST endpoint on /api/v1.
Five things a rep or a manager does with it
The 5 pm recap
One prompt. The agent pulls the day grouped by prospect and writes who was reached, what they said, the objections in their words, what was promised, and one move per prospect toward a meeting. The prompt and the structure are in the recap guide linked below.
The callback sweep
Every prospect who said “call me Thursday” this week becomes a task for Thursday morning, created from the transcript rather than from memory. The agent reads the disposition, not the rep’s recollection of it.
The objection digest
Across a week of transcripts: which objection came up most, in the prospects’ own phrasing, and which rep handled it best. This is the reading a manager never has time to do by opening forty recordings.
List building from a pasted spreadsheet
Paste fifty rows into the chat. bulk_create_leads imports them, tag_lead marks the batch, enroll_lead_in_sequence starts the sequence. Three tool calls, no CSV mapping screen.
The team recap for the admin
The same recap with scope set to team returns every rep’s day, the rep named on each prospect, so a floor of eight gets one page instead of eight.
How it is built, because it decides what you can trust
Every tool is a thin wrapper over the REST API at /api/v1. The agent can do nothing that a curl with your API key could not, and the same code serves both, so the two never disagree. The server is stateless: each request carries the key, is checked against the plan, and runs as that user. There is no separate permission model for agents, because the agent is you.
claude.ai and Claude Desktop connect through OAuth, which ends in an ordinary API key created in your account and listed under Settings → API Keys as “Claude (MCP connector)”. Claude Code and Cursor send a key you made yourself. Deleting the key disconnects the client on its next request. Tokens do not expire on their own; revocation is the delete.
Where it is the wrong tool
If your system of record is a large CRM, DialSheet's server covers the calling layer and the calling CRM; there is no native sync to Salesforce or HubSpot, only CSV and the API, so an agent that must update opportunities there would pair two connectors. If you want coaching while the call is live, this works after the call, from the recording. If your reps are on the Free tier, there is no API or server to connect. And if you need a push the moment a call ends, there are no webhooks on call completion; the agent reads when asked, or on a schedule you set on its side, and a script polls with a since parameter.
Pick it if you are a team of one to thirty on your own Twilio who wants the calls readable by the assistant you already pay for, without a conversation-intelligence contract that starts around $100 a user a month on an annual term. That is the reader this server was built for, and the reader it fits.
What it costs
The tokens an agent spends reading transcripts are billed by Claude or Cursor, not by us. A day of thirty answered calls is a few thousand words of transcript.
Common questions
Which power dialers have an MCP server?
DialSheet has one, hosted at https://www.dialsheet.io/api/mcp, with 30 tools covering calls with transcripts and AI summaries, a daily recap, leads, notes, tasks, tags, deals, sequences, email and SMS. It is part of Pro ($29/month, then $19 per seat). We have not found another dialer in the $0 to $50 a seat range that publishes one; the CRMs that do (the large ones) cover their records, not the calling layer, and you would pair the two.
What is an MCP server, in one paragraph?
MCP (Model Context Protocol) is an open standard Anthropic published in November 2024 that lets an AI client such as Claude call tools on a server. Instead of you copying a transcript into a chat window, the model asks the server for the calls it needs, reads them, and calls other tools to act on what it found. The server decides what is exposed; the client decides when to use it; your login decides what it may see.
What can an agent do with a dialer that it cannot do with a CRM export?
Act, and act on today. An export is a snapshot you hand over; an MCP connection lets the agent read this afternoon’s transcripts, create the callback task for the prospect who asked for Thursday, move the deal that verbally closed, and draft the follow-up from what was actually said, then do it again tomorrow without you touching a file.
Does it cost extra?
No. The MCP server and the REST API are included in Pro at $29/month, then $19 per seat, on top of the Twilio usage you already pay (~$0.014/min for a US minute, ~$1.15/month for a number). The tokens the agent spends are on your Claude or Cursor plan. Free-tier accounts do not have the API or the server.
Can the agent see the whole team’s calls?
A rep’s connection sees the rep’s own calls. An organisation admin can ask for a team recap, which returns every member’s calls for the day with the rep named on each prospect. There is no team-wide call list beyond that on purpose: the key belongs to a person and behaves like their login.
Is the MCP server the same thing as the AI call summaries?
They are two layers. Call Intelligence, in Pro, records, transcribes and summarises each answered call and stores the result on the call. The MCP server is the door through which your own agent reads those results and everything else in the account. You can have the summaries without ever connecting an agent; the agent has nothing to read without them.
Tool list and prices as of September 10, 2026. We build DialSheet, so weigh this page accordingly; the tool table is checkable in a minute by connecting and asking the client what it can do.