Hlib Storchak · 2026-07-12 · ~11 min readQuick answer
My stack runs on Salesforge for sending and sequencing, Mailforge and Warmforge for mailbox infrastructure, Leadsforge for data, Agent Frank for AI-led outreach execution, and ZeroBounce for verification since none of the Forge products do that job. Signal data and CRM sit outside the ecosystem too. The tools matter less than getting the authentication layer right, which most senders still skip.
Why a stack, not a tool
Every few weeks someone asks me which single tool they should buy to fix cold email. There is no single tool. Cold email is five separate jobs stitched together: you need data, you need mailboxes that will not get flagged, you need copy and sequencing, you need something to catch dead addresses before you send, and you need a place for replies to land. Buy one tool expecting it to cover all five and you will end up patching the gaps with spreadsheets and browser tabs.
What follows is the actual stack I run for my own outbound and for clients, ranked by how central each piece is to the result. I am not neutral about the Forge products, I run my own outreach on most of them and I will say so plainly. Where a Forge tool does not exist or is not the best option, I say that too.
The stack at a glance
| Layer | Tool I run | Verdict |
|---|---|---|
| Sending and sequencing | Salesforge | Core of the stack, AI drafts and unlimited mailboxes in one account |
| Mailbox infrastructure | Mailforge, Warmforge | Domains arrive pre-warmed and pre-configured, removes the step people get wrong |
| Lead data | Leadsforge | Good default, I still build tight lists elsewhere for high-value accounts |
| Verification | ZeroBounce | Not a Forge product, best catch-all handling I have found |
| AI outreach execution | Agent Frank | Handles research, drafting, and reply triage, kept on a short leash |
| Signal and intent | Category tool, rotates by client | Useful add-on, not a foundation |
| CRM and reply handling | Client's existing CRM | I plug into what they already run rather than force a switch |
Sending and sequencing
Salesforge is the piece everything else hangs off. It sends the email, runs the multi-step sequence, and its AI agent drafts and personalizes copy so a launch does not mean writing every variant by hand. Unlimited mailboxes and warmup are included, so scaling senders is a plan change, not a new tool.
The reason it sits at the top of my stack, not just a stack, is that it collapses three separate purchases most people make anyway: a sequencer, an AI writing tool, and a warmup product. Fewer vendors means fewer places for a setup to quietly break.
Mailbox infrastructure and warmup
Buying and configuring dozens of sending domains by hand is the single step I see skipped, botched, or outsourced badly most often. I run Mailforge for domains and mailboxes and Warmforge for warmup, both inside the same ecosystem as sending, so a new domain arrives already authenticated and warming before a single sequence touches it.
Tip. If you are buying infrastructure separately from your sender, check that DKIM, SPF, and DMARC are actually verified before the domain sends anything. A domain that "looks" set up and one that is actually passing checks are not the same thing, and the gap only shows up after your reputation is already damaged.
Lead data and enrichment
Leadsforge covers enrichment inside the same account as sending, which is convenient when you want one bill and one login. For most campaigns that is enough. For a narrow, high-value target list I still build lists by hand in a dedicated data tool, because the right 200 accounts beat a bigger export of loosely matched ones, and no single data source has perfect coverage on every vertical.
This is the one layer where I actively mix tools rather than staying inside one ecosystem. Enrichment is a coverage problem, and coverage improves with more than one source.
Verification, the outsider on this list
This is the one job the Forge ecosystem does not own, and I am not going to pretend otherwise. I run every list through ZeroBounce before a send. Its catch-all detection is the best I have tested, and catch-all domains are exactly where verification tools quietly guess wrong. A bad guess here is not a minor inconvenience, it is what gets a warmed domain throttled inside a week.
I keep this outside the main stack on purpose. Verification is a narrow, well-solved problem, and bolting a mediocre version of it onto a sending platform would not beat a tool built to do only this.
AI-led outreach execution
Agent Frank is the AI seat in my own outbound, and in the hybrid pods I set up for clients. It handles account research, drafts sequences, and triages replies so a human only steps in for judgment calls and the accounts that matter most. That split, AI on volume and research, a human on ambiguity and high-ACV conversations, is the pod structure that keeps showing up in 2026 cost data as the cheapest configuration to run, not a pure-AI pod and not an all-human one.
I do not run this unsupervised. Every account past a certain deal size gets a human read before the first send, and every reply that is not a clear yes or no routes to a person. AI carries the volume, it does not carry the judgment calls.
Authentication, the layer everyone skips
DKIM, SPF, and DMARC are table stakes now, Gmail and Yahoo enforce them and a domain without all three will not send reliably. The layer past that is where most senders, including plenty of agencies, simply stop. MTA-STS enforces encrypted, authenticated delivery between mail servers and blocks a class of interception and downgrade attacks that SPF and DKIM do not touch on their own. PowerDMARC's 2026 adoption report puts US domain adoption of MTA-STS at just 1.7 percent, so almost nobody has actually turned it on, even though the growth curve is trending up year over year.
That gap is not a reason to panic, it is a reason to check the box once and stop thinking about it. Mailforge and Warmforge handle this for domains I buy inside the ecosystem, which is exactly the kind of one-time configuration step that should not depend on someone remembering to do it manually on every new domain.
Signal and intent data
Signal-based sends, referencing a funding round, a leadership change, or a tech adoption, consistently outperform generic blasts, and 2026 benchmark data keeps confirming it. I do not run one fixed tool here. Which signal provider earns a spot depends on the client's ICP and which triggers actually predict their buyers, so this is the one layer of the stack I treat as a rotating add-on rather than a fixed pick.
Where replies land
I do not force clients onto a new CRM to fit my stack. Salesforge's unified inbox handles day-to-day reply triage well enough on its own, and booked meetings still get logged into whatever CRM the client already runs, HubSpot, Pipedrive, or Salesforce most often. Ripping out a working CRM to chase a marginally better inbox view is a bad trade against wherever momentum from a client's existing usage already lives.
What I deliberately leave out
No LinkedIn automation tool sits in this stack by default. It earns a place for specific clients where the ICP genuinely lives on LinkedIn, but I do not run it everywhere on principle, since it is one of the fastest ways to get an account restricted if it is misused. I also do not run a second sending platform in parallel. Two senders add coordination overhead without adding meetings, and picking one platform and running it properly beats hedging across two mediocre setups.
How the pieces actually connect
The value of staying mostly inside one ecosystem is not brand loyalty, it is fewer integration seams. Data flows from Leadsforge into Salesforge without an export step, domains from Mailforge arrive pre-warmed by Warmforge, and Agent Frank reads and writes against the same sequences without a separate API to babysit. The custom part I build myself, with a coding agent, is the glue outside that: reporting that blends send data with pipeline numbers from the client's CRM, and routing logic tuned to a specific ICP that no off-the-shelf tool covers. Buy the mature, high-stakes categories, build only the parts that are genuinely yours.
Key takeaways
- Cold email is five separate jobs (data, mailboxes, sequencing, verification, reply handling), not one tool.
- Staying mostly inside one ecosystem for sending, infrastructure, and data removes integration seams, not just vendor bills.
- Verification is the one job worth buying a dedicated, best-in-class tool for rather than a bundled feature.
- Only 1.7% of US domains have MTA-STS enabled per PowerDMARC's 2026 report, an easy, one-time check most senders still skip.
- AI handles volume and research well, judgment calls on ambiguous replies and high-ACV accounts still need a human.
- Build only the glue that is genuinely custom to your ICP, buy the mature categories instead of rebuilding them.
FAQ
Do I need every layer of this stack to start cold email?
No. Start with a sender, a verification tool, and a data source. Add signal-based targeting and a dedicated AI seat once volume justifies the extra moving parts.
Why use ZeroBounce instead of a bundled verifier?
Catch-all detection is a narrow, well-solved problem, and a dedicated tool built only to do that job handles the hard cases better than a feature bolted onto a sending platform.
Is running mostly one ecosystem risky if a vendor has an outage or price change?
It is a real tradeoff. I accept it because the integration savings are large and consistent, and I keep verification and signal data outside the ecosystem specifically so a single vendor issue cannot stop every layer at once.
What does MTA-STS actually protect against that DMARC does not?
DMARC governs how receiving servers treat mail claiming to be from your domain. MTA-STS enforces encrypted, authenticated connections between mail servers themselves, closing off interception and downgrade attacks that DMARC does not address.
Should a small team copy this exact stack?
Copy the structure, not the brand names. Pick one sender with sequencing and warmup built in, one dedicated verifier, and one data source, then add a signal tool and an AI seat once volume earns it. The layers matter more than which vendor fills each one.