One Agent, Many Tools: How We Wire GTM Work Through MCPs

We get asked a lot whether we are 'going all-in on MCPs.' The short answer is yes — with caveats. One agent with every tool in the stack plugged into it is powerful. One agent with every tool and no constraints is a liability with a nice UI.

Model Context Protocol (MCP) is useful because it turns tools into a standard surface the agent can call. It does not replace product design, permissions, or judgment. It just removes the tab-hopping tax.

The stack we actually plug in

In a typical go-to-market agent, tools fall into three buckets:

• Data & enrichment — crawl, enrich, and qualify accounts and contacts (Firecrawl, Apollo-style enrichment, lead intel, site scrape pipelines).
• Outreach & CRM — HubSpot, sequencing tools, Slack, Notion, call transcripts, task systems.
• Build & automate — n8n, GitHub, scrapers, browser sandboxes, vector search, and the database that holds durable state.

Same tools teams already love — they just stop opening them one by one. One prompt replaces one tab per tool. That only works if context lives somewhere durable.

Context before action

Our ICP definitions, playbooks, CRM notes, and call transcripts sit in one repo the agent reads before it runs anything. Without that, every tool call is a coin flip dressed up as automation.

A typical run looks like this:

• Pull companies that hit the pricing page this week
• Enrich decision makers
• Check them against HubSpot
• Draft the sequence in our voice
• Push it to the outbound tool — only after policy checks pass

Last year that took five tools and an afternoon. Now it can be one line — if the guardrails and CRM write policies are already in place. The coffee is optional. The checks are not.

What breaks when you skip the boring parts

• Duplicate contacts and polluted CRM stages
• Outreach that ignores unsubscribe and compliance state
• Agents that 'helpfully' overwrite human notes
• Cost spikes from enrichment loops with no rate limits
• Silent failures when one MCP server is down and the agent invents a workaround

MCPs connect the tools. The repo gives them context. Permissions, evals, and human approval gates decide whether the run is safe to ship. That combination is the product — not the connector count.