Here is something most outbound teams I talk to have not caught up to yet. I can hand a company domain to Claude Code and two minutes later get back a full discovery brief: revenue, decision makers, tech stack, top five competitors, recent news, and a tailored set of questions to ask on the call. No spreadsheets. No tab juggling. No SDR on my team spending forty-five minutes preparing for a thirty-minute meeting.
That is one workflow. I run three more like it: event attendee scraping, a daily hot leads digest pushed to Slack, and post-call admin that updates the CRM and drafts the follow-up email for me. Below is exactly how I built each one, and how the whole stack snaps together when your sending platform exposes its actions to Claude Code as native tools.
Claude Code is Anthropic's command line AI coding agent. I describe what I want in plain English and it writes the code, runs it with my approval, and returns the result. I use it to build small internal tools, scrape data, and move information between systems without writing code myself.
Claude co-work is a desktop AI agent. I give it access to my laptop and it operates my CRM, Slack, Gmail, and calendar on a schedule. It is what I point at recurring work that needs to run every morning without anyone clicking buttons.
Simple rule I follow: Claude Code for things I want to build once, Claude co-work for things I want to run every day. The four workflows below split cleanly across both. The first two are Claude Code builds. The last two are co-work agents.
Most SDRs and AEs I know still do pre-call research by hand. Forty-five minutes across LinkedIn, the company site, Crunchbase, and G2, and they still do not know how to run the call. Here is how I rebuilt mine. Total build time: about an hour.
Every Claude Code project starts with a CLAUDE.md file in the root. I treat it like an onboarding doc for a new SDR: a paragraph on what I sell, my ICP, my top three differentiators, and tone rules for the brief. Without this file the output is generic. With it, the brief reads like one of my AEs wrote it.
Agent 1 (company): takes a domain, returns revenue band, headcount, named decision-makers, tech stack, and hiring signals. Agent 2 (market): top five competitors, recent news, growth trajectory. Agent 3 (discovery strategy): takes the output from agents 1 and 2 plus my CLAUDE.md context and produces the final one-page brief with three likely pains and five tailored discovery questions.
When a meeting is booked in HubSpot, a webhook fires my Claude Code command with the prospect's domain and posts the brief into the deal's Slack thread. Latency from booking to brief in Slack: under two minutes. No HubSpot? Run the command manually and paste the output. Same build, different trigger.
Inside Salesforge, I trigger this whenever a new meeting books, so the brief is waiting in Slack before my AE opens the calendar invite. Before: 30-45 minutes per call. After: under 2 minutes.
Events are expensive. A conference in San Francisco can run me three thousand dollars before I shake a hand, and the old process for booking meetings was four to six hours of screenshotting and tab juggling for a 200-person attendee list. With Claude Code and Firecrawl, that collapses to minutes.
Firecrawl handles the messy parts of scraping: JavaScript rendering, login flows, structured extraction. I add it to Claude Code as an MCP server, which means Claude Code sees Firecrawl's scrape and extract functions as native tools. No custom code.
I do not tell Claude Code how to scrape. I tell it what shape I want the data in. My prompt looks like: "Extract every attendee from this URL into a CSV with full name, title, company, and LinkedIn URL. Skip anyone whose title is not founder, CEO, head of, VP, or director." Claude Code figures out pagination and selectors on its own.
Most paid events give me login credentials to a matchmaking app. I pass them to Claude Code in the prompt, Firecrawl handles the session, and the scrape runs through the auth wall. 2FA is the only blocker, and most event platforms do not enforce it. If you would rather not pass credentials, log in yourself and ask Claude Code to extract from a saved HTML file instead.
The Firecrawl CSV has names and titles but rarely verified emails. I push it through Leadsforge for waterfall enrichment, then load it straight into a Salesforge campaign. A two-day sequence against one Ecom conference list booked five meetings for me in under 30 minutes of active work. Before: 4-6 hours per event. After: 20-30 minutes.
Worth flagging: a 200-person event list will burn through mailbox capacity fast if your sending tool charges per inbox. Salesforge makes this practical for me because unlimited mailboxes are included on every plan, and Claude Code can drop the enriched list into a campaign through the Salesforge MCP.
Every outbound team has the same pipeline leak. You promise to follow up in two weeks. Two weeks pass. You forget. The lead goes cold. My fix is a Claude co-work agent that runs every morning before my team starts work.
The agent needs read-only access to two systems: my CRM (HubSpot) and the call transcripts attached to the deals (Fathom, Gong, whichever you use). Read-only is the point. This workflow surfaces leads, it does not act on them.
The agent needs scoring rules in a markdown file, otherwise it surfaces everything. Mine: stage past initial demo but before closed-lost, last touch between 7-30 days ago, at least one positive sentiment marker in the most recent transcript, and deal size above my median ACV. Boring on purpose. The goal is consistency, not cleverness.
Claude co-work runs the workflow every weekday at 7:30 a.m. It scans the pipeline, ranks the top 15 by combined recency and signal, and writes a Slack message where each card has the company name (deep-linked to HubSpot), days since last contact, a one-sentence summary of the last call, and a priority flag on the top three. AE clicks the company name, lands in the CRM, makes the call. No dashboards, no pipeline reviews.
For higher-volume teams like mine, the same agent can trigger an AI voice follow-up on leads that already consented to calls, so the human only gets pulled in when the prospect is back in active conversation.
Every AE on my team used to lose 30 minutes after every call updating the CRM, writing notes, and drafting follow-ups. Five calls a day means 2.5 hours of admin during peak energy hours, when I want them closing instead.
Claude co-work needs the transcript source (Fathom, Gong, Granola) and write access to the CRM, but I scope CRM permissions tightly. The agent updates only whitelisted fields: call summary, next step, deal stage, follow-up date. Nothing else. Critical rule: never advance a deal stage without explicit verbal agreement in the transcript. False positives in the pipeline are worse than no automation.
The agent drafts the follow-up in the AE's voice, including the next step and any resources promised on the call. The draft drops into a Slack DM. AE reviews for two minutes, edits, hits send. Human in the loop on every external touch.
Claude co-work watches my meeting recorder. When a transcript becomes available, the agent fires: extract data, update fields, draft email, ping the AE in Slack. End-to-end latency from call ending to draft ready is under five minutes. Before: 2.5 hours of admin per day per AE. After: 10 minutes of review.
If your follow-up volume is high enough that even review becomes a bottleneck, Agent Frank can take over the whole outbound loop as my AI SDR, so the only thing left on my plate is closing.
All four workflows above share the same final step: at some point, the leads I researched, scraped, prioritized, or followed up on need to actually get sent. That is where the choice of sending platform stops being neutral. Salesforge is the only outbound platform I have found that ships an official MCP (Model Context Protocol) server, which means Claude Code operates the entire Forge stack as native tools without scripts or API wiring.
MCP is the standard Anthropic built for giving AI agents native, tool-level access to external systems. I describe a job in plain English. Claude Code picks the right MCP tool. I approve the action. The tool runs. No tab switching, no copy-pasting credentials, no integrations to maintain.
This is what makes the four workflows above stop being four separate experiments and start being one operating model. The research agent finds the company, the scraper builds the list, the digest surfaces the warm leads, the post-call agent updates the record, and Salesforge handles every send in between, all from the same Claude Code session.
In practice, it covers the work I used to do manually inside the UI:
One note: Mailforge does not ship a standalone MCP surface. It connects natively through Salesforge, so any Mailforge mailboxes already attached to a workspace are reachable through the Salesforge MCP tools.
The practical impact: setting up sending names, profile pictures, warmup schedules, campaign drafts, and lead uploads used to be an afternoon of clicking through the UI for every new client my team onboarded. With the Salesforge MCP, the same work is now a single Claude Code session. Paste the onboarding details, approve each action, the stack is live.
Quick reference for the tools I use to power each workflow above.
Claude Code is Anthropic's command line AI coding agent. It runs in your terminal, builds small internal tools, scrapes data, updates your CRM, and operates the Forge stack through the Salesforge MCP. ChatGPT is a chat interface. Claude Code is closer to an on-demand operator that does the work for you.
No. You describe what you want in plain English and Claude Code handles the execution. The main skill is writing clear instructions and giving it good context through a CLAUDE.md file with your ICP, tone rules, and any product docs.
Start small. Use a sandbox account for experiments, give it read-only access first, and scale only once a workflow is stable. Treat it like any new employee onboarding to sensitive systems.
Claude co-work is a desktop agent that operates your apps on a schedule (CRM, Slack, Gmail). Use Claude Code to build something once. Use co-work for recurring tasks that need to run every day on their own.
The Salesforge MCP is an official Model Context Protocol server that exposes Salesforge and Forge stack actions to Claude Code as native tools: launching campaigns, adding leads, rotating sending accounts, running Warmforge checks, and building Leadsforge lists. You describe the job, Claude Code picks the right tool, you approve. No scripts.
A pre-call research agent takes about an hour. Event scraping with Firecrawl takes minutes once you know the target URL. Post-call admin and hot lead digests take a week to get right because they depend on your CRM structure and scoring rules.
.png)
