Running outbound from your terminal: the MCP setup, line by line
Your MCP config is a few lines of JSON. What each line buys you, what to connect first, and the mistakes that burn domains and credits.
Outbound used to mean buying a seat in a sales engagement platform and living inside its UI. The MCP route is different: your agent does the work, from the terminal you already have open, through one server that carries sourcing, sending, calls, and content. This post is the setup, line by line, including the parts people get wrong.
The config itself
One entry in your MCP config: a server name, the command or URL that starts it, and your API key. That's the whole integration. No OAuth dance per tool, no CSV imports between systems. When the server starts, your agent discovers what it can do the same way it discovers any tool.
Keep the key in your environment, not in the config file you commit. The config is code; the credentials are secrets.
What the server carries
Sourcing: company and contact search across 100+ providers, with filters for size, industry, funding, and geography, and dedupe against your CRM before anything else happens. Outbound: mailbox infrastructure on managed domains, so deliverability is the server's problem, not a warmup script on your laptop. Calls: an outbound caller that can place, qualify, and book. Content: first-touch drafts written from your positioning doc, not from a template library.
The mistakes that cost real money
- Sending from your primary domain. Cold email on your company domain puts your deliverability at stake with every campaign. Managed domains exist so a bad batch costs you a domain, not your inbox.
- No dedupe before send. Source first, check the CRM second, send third. Any tool that skips the middle step will email your existing customers.
- Unbounded runs. "Find everyone in fintech" is an unbounded ask. Cap the run: twenty companies, one segment, one message. Credits buy work; scope decides how much work.
- Reviewing nothing. The first ten drafts tell you whether the voice is right. Read them. Then the review gate can loosen.
How a run actually goes
A typical session: describe the segment and the cap, let the agent source and dedupe, read the research summaries, approve the drafts, and let the sends run on schedule. Replies and call outcomes write back to the CRM as resolved records. The whole loop runs from the terminal; the only UI you touch is your CRM, when you want the history.
What this replaces
The seat-based stack: a list builder, an engagement platform, a dialer, an AI writer, and the glue between them. Each one had a login, a billing line, and an export button. One MCP server with one credit pool replaces the glue and most of the logins. What it can't replace is taste: the positioning doc and the review gate are still yours.
The RevPeak MCP server connects with one command, and outbound plays run on plans from $34 a month.