How to run sales prospecting and outbound from Claude Code

What a GTM team can do from Claude Code

Claude Code is a coding assistant, but most of what it does is read, plan and run commands. That is also most of outbound: decide who to target, find them, learn enough to write something relevant, send it on the right channel, and keep the CRM current when people answer. Once Claude Code can call a GTM workspace, each of those becomes a sentence you type instead of a tab you open.

The work stays the same as your BDRs do it today. What changes is who does the clicking. You describe the outcome ("200 Heads of RevOps at US SaaS companies, with verified emails, in a sequence that starts on LinkedIn"), Claude breaks it into steps, shows you what each step returns and costs, and asks before anything leaves the building.

• Account research: hiring, funding, news and tech stack for a target account, with sources, written onto the company record.
• Building lists: people and company searches by title, industry, headcount and location, saved into a table.
• Enriching lists: work emails and phone numbers through waterfalls, plus AI columns that score fit or write a first line per row.
• Writing sequences: email, LinkedIn and WhatsApp steps in one sequence, with branches and A/B versions of a step.
• LinkedIn steps: profile visits, invites, messages and InMail from your connected account, sent in the cloud.
• Updating the CRM: companies, people and deals, with replies landing on each person's timeline.
• Reading replies: one inbox across email, LinkedIn and WhatsApp, with AI triage and drafted answers to review.

How Claude Code reaches your GTM tools: MCP servers, skills and plugins

Claude Code gets new abilities in three ways, all documented by Anthropic. An MCP server gives it tools: you add a remote one with claude mcp add --transport http <name> <url>, and sign in by running /mcp inside Claude Code and finishing the login in your browser. A server can be added for you in one project (local, the default), for every project on your machine (user), or for the whole team through a .mcp.json file in the repository that you commit (project).

A skill is a SKILL.md file of instructions. Personal skills live in ~/.claude/skills, project skills in .claude/skills, and Claude loads one when its description matches what you asked, or when you type /skill-name. A plugin packages skills, agents, hooks and MCP servers as one unit you install from a marketplace.

Anthropic's MCP guide also says to verify you trust each server before connecting it, because servers that fetch outside content can expose you to prompt injection. For sales work that matters: a server that reads inbound replies or web pages is reading text written by strangers, so actions with consequences (a send, a paid lookup, a CRM change) should need your approval, whoever's server you use.

Sources: Claude Code docs: Connect Claude Code to tools via MCP, Claude Code docs: Extend Claude with skills, Claude Code docs: Plugins overview

What OXYGEN gives Claude Code to work with

OXYGEN is one GTM workspace that Claude Code operates through a CLI or an MCP server. It holds a CRM (companies, people, deals and objects you define, a deals pipeline and an activity timeline), tables you edit like a spreadsheet with enrichment and AI columns, email waterfalls across 17 providers and phone waterfalls across 13, charged only when one finds the value, and sequences that mix email, LinkedIn and WhatsApp.

Because those parts share one record per lead, the joins you would otherwise write yourself are already there. A reply on LinkedIn stops the email steps for that lead. The replier is filed as a person on the CRM. A lead you enriched last month is not paid for twice. And every run Claude starts shows up in the web app with its cost, results and a link, so a teammate who never opens a terminal can still see what happened.

The web app, the CLI and the MCP server call the same contract, so state Claude creates from the terminal is the same table, sequence or record your team opens in the browser.

Approvals before sends and paid steps, costs in credits

Previews and dry runs make no provider calls and spend nothing. A paid provider call, an AI or tool column run, an enrichment run, a CRM write, an outbound enrollment or a LinkedIn action refuses to execute until it is approved, and anything that carries a price also needs a credit ceiling. From the CLI that is --approved with --max-credits; over MCP it is approved: true with max_credits.

The ceiling is a cap, not a charge: unused reserved credits are released. When a gate stops a command, it exits with code 7, spends nothing and names the next step, which gives Claude Code a clean place to stop and ask you. Afterwards, oxygen billing usage shows what each run actually cost. One exception to know: a sequence's own credit ceiling is optional and off unless you set it.

Why not Claude Code with Apollo, lemlist and HeyReach directly?

You can, and for some teams it is the right call. Apollo runs a hosted MCP server at mcp.apollo.io/mcp that Claude Code adds with one command, with people and company search, enrichment and sequences; its enrichment actions use Apollo credits. lemlist's server at app.lemlist.com/mcp manages campaigns, leads, lead sourcing and email finding over OAuth. Instantly exposes its V2 API as MCP tools at mcp.instantly.ai with an API key. Clay MCP brings its providers, research agents and Functions into Claude, ChatGPT and Codex. HeyReach ships an MCP server for LinkedIn campaigns, leads and replies.

What you take on is the glue between them. Each server has its own lead record, its own auth, its own credits and its own idea of a campaign. To run one motion across them, you or Claude has to keep IDs in sync, stop LinkedIn steps when someone answers an email, suppress customers and existing deals, write replies back to the CRM, add up spend across four bills, and keep your approval rules in prompts and skills that you maintain as each vendor changes its tools.

OXYGEN puts those parts in one workspace instead. If you already pay for one of these tools, they can sit side by side: OXYGEN pushes leads into Instantly, Smartlead, lemlist and HeyReach campaigns, and searches Apollo's database on your own Apollo key.

Sources: Apollo docs: Apollo MCP, lemlist developer docs: MCP setup, Instantly developer docs: MCP introduction, Clay guide: Clay MCP, HeyReach: HeyReach MCP

CLI or MCP server: which to use in Claude Code

In Claude Code the oxygen CLI is the primary path. It has the full command surface, returns one JSON envelope on every command with a web_url to open the result, and uses exit codes Claude can branch on without parsing text. oxygen skills install --agents claude-code writes one skill per OXYGEN feature (tables, enrichment, sourcing, the sequencer, onboarding) into ~/.claude/skills, so Claude knows the right commands before it starts.

The hosted MCP server at https://oxygen-agent.com/mcp is optional and additive. It signs in with OAuth, so no key goes into a prompt. Claude Code does not render OXYGEN's inline table widgets, so open the web_url in each result to look at the grid. For large CSV files, import them from disk with oxygen tables import rather than pasting them into a prompt.

Claude Code for marketing teams

The same workspace covers the marketing side of GTM. Claude can schedule LinkedIn and X posts that you approve one by one before they go out, read LinkedIn and X analytics, and pull free signals from your connected LinkedIn accounts: profile viewers, new followers and people who reacted to a post. Those signals can feed a table or a sequence, so the people who engaged with this week's post become next week's list, with research and an opening line written per row.

What OXYGEN does not do here

• No contact database of its own: searches run through data providers on credits, or through Apollo's database on your own Apollo key.
• No Salesforce sync. HubSpot syncs both ways, Attio one way, and Pipedrive is an import.
• Claude Code does not show OXYGEN's inline table widgets; open the web_url in each result instead.
• LinkedIn automation carries account risk under LinkedIn's terms. Daily and weekly limits per account lower it; nothing removes it.
• No SMS step, and call tasks are in beta.
• Claude Code itself is billed by Anthropic. OXYGEN bills its own work in credits.