A GTM agent that runs outbound, enriches CRM, and costs $2 per run

Most GTM AI content falls into one of two buckets: prompt tips for ChatGPT, or $500/month tool comparisons nobody can actually act on. This is neither.

This is a breakdown of a working system I built and run: a four-layer Claude Code agent stack that handles outbound research, sequence drafting, CRM enrichment, and ICP scoring. No Clay. No Apollo enrichment credits. No n8n flows that break every three weeks.

Just Claude Code, a handful of MCP tools, a well-structured CLAUDE.md (grab the starter template), and five operator skills wired together.

Why Claude Code instead of Clay

Clay is excellent at what it does: structured enrichment pipelines with a visual interface and a credit model per row. I’ve used it. It’s genuinely useful for teams that need enrichment at scale without engineering resources.

But it has limits that matter for the way I work:

• You’re buying enrichment credits, not intelligence. Clay routes data through third-party providers: Clearbit, Apollo, LinkedIn scrapers: and bills per row. The LLM reasoning layer is thin.
• The visual pipeline is powerful but brittle. Complex logic becomes hard to inspect, version, or hand off.
• It doesn’t write. You can pull data into a variable and inject it into a template, but the actual reasoning about why this prospect matters for this account at this moment isn’t there.

Claude Code: specifically the agent loop via the SDK, with MCP tools wired in: lets me build something different: a system that reads context, reasons about fit, writes copy, and updates records, all in one pass. The intelligence is the product, not the infrastructure.

The four-layer stack

The system has four components that work together:

Layer 1: CLAUDE.md: the operator brief

Every Claude Code agent session starts by reading CLAUDE.md. This is not documentation. It’s the briefing file: the equivalent of what you’d give a junior SDR on their first day, except it’s read by the model on every run.

Mine includes:

• ICP definition (firmographic, technographic, behavioral signals)
• Tone and persona for outreach (direct, no buzzwords, operator-to-operator)
• What a qualified lead looks like vs. a time-waster
• How to handle edge cases (no LinkedIn, foreign markets, inbound vs. outbound)
• Hard rules (never mention competitors by name, never use “I hope this finds you well”)

The CLAUDE.md is the most important file in the system. It’s where operator judgment gets encoded so the model doesn’t have to guess.

Layer 2: MCP tools: the data access layer

I run four MCP servers in the stack:

Web search (via Brave MCP or Exa MCP): for real-time company research, funding news, job postings, recent content. This replaces the “waterfall enrichment” step in Clay.

Browser automation (via Puppeteer MCP): for pages that block crawlers. LinkedIn company pages, some pricing pages, Glassdoor signals about team size and burn rate.

Memory / file system (via the Claude filesystem MCP): for reading and writing client files: the CRM export CSV, the sequence templates, the running log of contacted accounts.

Custom scoring MCP (built in a weekend with the MCP SDK): takes a structured JSON object (company data + ICP criteria) and returns a fit score with reasoning. This was the hardest to build and the most valuable: it gives the agent a consistent scoring rubric across runs.

Layer 3: Client memory files

Each client gets a folder:

Layer 4: The five Skills

Skills are named sub-tasks the agent can invoke. I have five:

A real run, step by step

Here’s what an actual session looks like when I’m working a new account list:

1. Drop a CSV of 50 company names into /clients/acme-corp/
2. Run: claude /enrich clients/acme-corp/new-accounts.csv
3. The agent loops through each account: web search → browser scrape if needed → score → write research file → append to enriched CSV
4. 20 minutes later: enriched CSV with fit scores, priority flags, and a research file for every account above 6
5. For the top 10: claude /draft [company] v2: one at a time, reviewing each draft before committing
6. Log sent sequences with /log

Total time for a 50-account list: about 45 minutes of unattended agent work, plus 30 minutes of my review. Equivalent Clay + copywriting workflow: 3–4 hours minimum, and the copy quality isn’t comparable.

What this costs

The cost model matters. Here’s a rough breakdown per 50-account enrichment run:

• Claude API (claude-sonnet-4 at current pricing): ~$2–4 per full run depending on research depth
• MCP tool calls: Brave Search API is ~$0.003/query; Puppeteer is self-hosted so compute only
• Custom scoring MCP: minimal, it’s just structured JSON in/out

Compare to Clay: $800–2,400/month for a plan that gives you meaningful enrichment credits plus LLM rows. For my volume and use case, the agent stack runs at roughly 5–10% of that cost, with better output quality on the reasoning and copy layers.

The tradeoff: it took time to build. Probably 20 hours across the CLAUDE.md, the MCP setup, the Skills, and the scoring tool. For a team without that capacity, Clay is the right answer. For an operator who can invest the build time, this pays back quickly.

The CLAUDE.md structure I use

Since this is the most reusable piece, here’s the template structure I’ve landed on:

The starter pack

I’m releasing the core of this as the Claude Code GTM Agent Starter Pack: the CLAUDE.md template, the Skills definitions, the folder structure, and the scoring MCP source.

It’s available on GitHub. It’s free. It’s designed for GTM operators who want to build this themselves rather than wait for a SaaS wrapper to charge $400/month for it.

The pack assumes:

• You have a Claude API key
• You’re comfortable running Claude Code in terminal
• You understand your own ICP well enough to fill in a CLAUDE.md

If those three things are true, you can have a working enrichment and outreach agent running in an afternoon.

What’s next

I’m adding two things to the stack:

CRM sync: a bidirectional MCP connector to HubSpot that writes enriched account data directly to contact/company records without manual CSV import.

Signal monitoring: a lightweight cron that runs /score weekly against a watchlist of accounts and surfaces re-engagement triggers (new funding, leadership change, job postings for relevant roles).

Both are in progress. I’ll document them here when they’re stable.

If you’re building something similar or want to see the starter pack in action, book a call. I’m happy to walk through the architecture.

Related: GTM AI Agent Architecture: the four-layer stack guide · B2B CRM as a Revenue Operations System: how to rebuild it right