Pack de Ventas y Outreach con IA

What's in this pack

This is the stack for someone who actually has to book the meeting — a founder doing their own outbound at pre-seed, a single AE at a 20-person Series A, a solo consultant filling next quarter's pipeline. Not a Twitter list of 40 AI sales startups. Every pick does one specific job in the outreach cycle: research the account, enrich the contact, draft something that doesn't read like a template, send and log it, and capture the meeting cleanly back into the CRM.

The pack is built around one principle: AI accelerates the prospecting, but the contact graph is non-negotiable. That's why the CRM pick is self-hosted (Twenty), not a vendor lock-in. That's why the email layer is an MCP (Google Workspace or AgenticMail) you actually control, not a wrapper SaaS. And that's why every step writes back to the same record — so three months later you can answer "why did we reach out to this account" without grepping six tools.

Install in this order (research → enrichment → draft → send/log → meeting capture → nurture)

1. Claude Code Agent: Competitive Analyst — start here. Before you write any email, you need to know what the prospect's competitive landscape looks like. This agent benchmarks direct and indirect competitors so your pitch lands on the right tension, not generic platitudes.
2. Claude Code Agent: Research Analyst — deep account research. Read the company's blog, their recent funding, who left and who joined, what their job listings hint at. Output is a one-page brief you actually read before drafting.
3. Firecrawl — Web Scraping API for AI Applications — the enrichment layer. LinkedIn URLs, company pages, recent press, pricing pages — Firecrawl pulls them into clean markdown that the agents above can ingest. This is the spine connecting "a list of names" to "a list of named situations."
4. Claude Code Agent: Sales Automator — the drafter. Once you have the brief, this agent writes the cold email, the follow-up, the case-study reference, the proposal template. It is shockingly good at not sounding like a sequence — provided you give it the briefs from steps 1-3, not just "write an email to acme corp."
5. Google Workspace MCP — Gmail/Calendar/Drive for Agents — the send and log layer. The agent drafts in step 4, then this MCP puts it into Gmail as a draft you review, sends, and logs the thread back into the contact record. No more copy-pasting between tabs.
6. AgenticMail — Email + SMS Infra for AI Agents (MCP) — the programmatic alternative when you need volume. For multi-channel sequences (email + SMS), AgenticMail exposes both via MCP so the agent can fan out without you wiring SMTP and Twilio by hand.
7. Twenty — Open-Source AI CRM (Salesforce Alternative) — the contact graph. Self-hosted, GraphQL-first, and designed so an agent can read and write records without paying Salesforce per seat. Owns the canonical source of truth.
8. Claude Code Agent: Salesforce Expert — for teams already on Salesforce or HubSpot (most are). This agent handles SOQL queries, custom object schemas, and the migration scripts when you eventually escape. Pair with Twenty if you're standing up new; keep on its own if you're managing a 50k-record legacy org.
9. Meetily — Privacy-First AI Meeting Assistant with Local Transcription — the meeting layer. Records the discovery call, transcribes locally (no cloud upload), extracts action items and objections, and ships a structured note back into the CRM contact. Your prospect's roadmap secrets stay on your laptop.
10. Mautic — Open Source Marketing Automation Platform — the nurture loop. Cold-replied-not-yet-ready prospects go into a Mautic drip. Open and click events flow back into Twenty so the sales agent knows when to re-engage. The boring part of outbound that compounds the rest.

How they fit together

   Target Account List (CSV / LinkedIn / signal)
            │
            ▼
      Firecrawl (3)
   (enrichment, web context)
            │
   ┌────────┴────────┐
   ▼                 ▼
 Competitive (1)   Research (2)
 Analyst           Analyst
   │                 │
   └────────┬────────┘
            ▼
     Account Brief (one page)
            │
            ▼
   Sales Automator (4)
   (drafts email + follow-ups)
            │
   ┌────────┴────────┐
   ▼                 ▼
 Workspace MCP (5)  AgenticMail (6)
 (review + send +   (programmatic
  log to thread)     multi-channel)
   │                 │
   └────────┬────────┘
            ▼
     ┌─── Twenty (7) ─── Salesforce (8)
     │   (contact graph, source of truth)
     │            ▲
     │            │
     ▼            │
   Meetily (9)    │
   (call → notes ─┘
    back into CRM)
     │
     ▼
   Mautic (10)
   (nurture not-yet-ready replies)

The critical join is Twenty (or Salesforce) as the contact graph + agents as read/write clients. As long as every step writes back to the same record, the agents stay accountable. Lose the graph and you're back to spreadsheet roulette.

Tradeoffs you'll hit

• Personalization vs volume — the Sales Automator can write 300 emails an hour. It will not write 300 good emails an hour. Hard cap: 30-50 outbound touches per day per rep, each with a real Research Analyst brief attached. Above that, you're a spam filter's training data.
• Self-hosted CRM vs Salesforce/HubSpot reality — Twenty is real and good, but if your team already lives in Salesforce, replacing it is a six-month project. Pragmatic move: keep Salesforce as system of record, use Twenty as a sandbox for the agent-driven workflows, and migrate later if it sticks.
• Local transcription vs cloud — Meetily runs Whisper on your laptop. It's slower than cloud, but discovery calls often contain pricing, roadmap, and competitive intel the prospect would not want on a vendor's servers. Pay the latency tax for the trust premium.
• MCP email vs traditional sequencer — Outreach.io and Apollo.io are mature for a reason. The MCP layer wins when you want an agent to read the inbox and react (a reply triggers a deep research pass). It loses on raw deliverability tooling. Use both if you can.

Common pitfalls

• Letting the agent send without review — every cold email goes through a human eye for the first 90 days. The agent drafts, you press send. After 90 days you'll know which patterns are safe to auto-send and which aren't.
• CRM as a graveyard — if the agent writes notes nobody reads, you have a worse CRM, not a better one. Set up a weekly Twenty/Salesforce digest of agent-written records that the sales lead reviews. Drift gets caught early.
• Enrichment hallucination — Firecrawl returns real markdown; the agents on top can still make up facts about what they read. Mitigation: cite every claim in the brief back to the URL it came from. No URL = the claim doesn't go in the email.
• Compliance forgotten — GDPR, CAN-SPAM, CASL, and state-level email laws apply to AI-generated outbound the same as human-generated. Suppression lists, unsubscribe links, and physical address footers are still mandatory. The agent will happily skip them; you can't.
• Meeting recording without consent — Meetily is local-only, but you still need explicit consent from every participant on the call. "This call is being recorded for our notes" at the top — every time. Not optional.