MCP for sales

# MCP for sales

MCP for sales is the application of the Model Context Protocol, an open standard published by Anthropic in late 2024, to the tools that sellers and revenue teams use day to day. The protocol gives an AI agent a single common interface to call out to external systems. For sales that means a CRM, an email sender, a prospect database, an enrichment provider, and signal sources can all be reached by the same agent without writing custom glue.

Before MCP, an agent integration was bespoke. Every model vendor had its own function-calling format, every tool had its own SDK, and every connection between them needed engineering work to maintain. After MCP, a tool ships one MCP server. Any compatible agent reads the server's skill list at startup and calls whatever capability it needs. The integration cost drops by roughly an order of magnitude.

Why it matters

Sales is one of the highest-coordination workflows in a typical company. A single outbound motion touches list building, enrichment, drafting, sending, reply triage, CRM updates, and handoffs to AEs. Each of those steps lives in a different tool. Each tool has its own API, its own permissions model, its own auth flow. Coordinating them is what burned out the SDR role.

An agent can in principle coordinate all of those steps. The blocker has not been model capability. The blocker has been integration plumbing. To run a real outbound motion the agent had to know how to authenticate against Apollo, Salesforce, Smartlead, Clay, LinkedIn, and the user's calendar. Building and maintaining that surface for each new agent runtime was a tax that killed most projects before they shipped.

MCP shifts the burden. Each tool ships its skills once. The agent reads them at startup. The user picks which skills to expose. The same prompt that worked yesterday with Claude Desktop works tomorrow with Cursor or a custom runtime, because the protocol is the contract.

The practical implication is that the buyer side of the market has reasonably strong leverage now. A sales tool that does not ship an MCP server cannot easily participate in agent-driven workflows. A buyer who is committing to that workflow can require MCP support in evaluation. This is reshaping the procurement criteria for outbound tooling in 2026.

Common pitfalls

The first pitfall is treating MCP as a marketing checkbox rather than a real capability. Many vendors have shipped MCP servers that wrap their existing REST API and expose every endpoint. That is technically MCP support but it is not useful to an agent. The agent needs skills, named capabilities with clear arguments and expected outputs, not a raw API surface. See agent skill manifest for the right way to expose this.

The second is over-privileged auth. The temptation when wiring up a sales agent is to give it the same permissions a senior rep has. That is a lot of latitude, and the agent can move quickly in directions you do not want. The right pattern is least-privilege scoping per skill, with destructive actions like bulk-delete or send-all requiring an extra confirmation step.

The third is ignoring rate limits and back-pressure. An agent that wants to be helpful will fan out parallel calls to enrichment providers, the sender, and the CRM all at once. If the MCP server does not surface the underlying rate limits clearly, the agent will exceed them silently and the run will fail halfway through.

The fourth is poor error surface. When a skill call fails, the agent needs a structured error it can reason about ("rate limit exceeded, retry in 60 seconds") rather than an opaque 500. Vendors that ship MCP without this hygiene cannot be trusted for production use.

How it shows up in agent-driven workflows

Coil does not currently ship an MCP server. Its agent surface today is an organization-scoped API, a public JSON CLI, capability manifests, and runtime-native skills for Codex, Claude Code, Hermes Agent, and OpenClaw. Those surfaces support agent-operated workflows without claiming protocol compatibility that has not shipped.

If Coil support supplies reviewed skill tarballs or a ClawHub folder plus SHA-256 checksums through a release handoff, verify those artifacts before installation. Store the organization API key in the runtime secret manager and start with coil agent-context --json.

This is what agent-driven outbound looks like in practice. The model does the reasoning. Coil's authenticated API or JSON CLI carries its calls today. The user reviews and approves the output.

Related concepts

Agent skill manifest is the discovery file that an MCP server publishes so agents know what capabilities are available. A skill manifest is to MCP what an OpenAPI spec is to REST, except aimed at being read by a language model.

Agent-driven outbound is the motion that MCP enables for sales teams. The protocol is the means, the motion is the end.

Lead enrichment is a frequent skill the agent will call through MCP, often against more than one provider to triangulate data quality.

Sales Navigator scraping is one of the more common skills a sales-focused MCP server will expose, with the caveat that LinkedIn's terms still apply regardless of the protocol used.

For more on Coil's current agent-operated architecture, see the agent-driven lead gen pillar.