The AI tools we use to run a B2B marketing system.

A stack report, not a listicle. Twenty-seven tools currently authorized on the corporate card, ranked internally by how many times they pay back their monthly cost. The ones we can't run without, the ones we quietly pay for, and the three we cancelled this quarter.

Tier A — Non-negotiable

Five tools we would not re-found the firm without.

01 · Claude

Our primary model for any task requiring long-context reasoning, writing in a specific voice, or navigating ambiguity. The 200k token context is the hidden feature. We feed entire source corpora in and ask for synthesis; nothing else handles that shape of task as well today.

02 · GPT-4 class (OpenAI)

Used for tasks that benefit from its faster edit-loop tempo and structured-output discipline. Specifically: routing, classification, short-form generation, and any place where we want tool-calling semantics that the OpenAI SDK handles cleanly.

03 · Postgres (Neon)

Not a model. The actual foundation of every agent we ship. Every run, every outcome, every piece of memory lives here. If we had to cut everything and keep one tool, we would keep Postgres.

04 · Claude Code

The terminal agent that writes most of our internal code. We use it for scaffolding new agents, refactoring prompts, and exploratory data work. It's the fastest way we've found to go from "idea" to "running job" in under an hour.

05 · Cursor

When code needs human judgement — design decisions, architecture — we work inside Cursor with the same models. It's the IDE level of the same pattern Claude Code is on the terminal.

Tier B — High leverage

Nine tools that are the difference between running a system and running spreadsheets.

Resend for every transactional email. Smartlead for cold-outbound sending and warm-up. Apollo for bulk enrichment. Neon as our Postgres host. Temporal Cloud for durable agent workflows. Vercel for every dashboard we ship. Linear because project management is not a problem worth rebuilding. HubSpot at almost every client as CRM. Clay for one-off enrichment flows that don't justify custom code.

Tier C — Useful

Ten more tools that earn their keep quietly. The fully-qualified list: Notion (docs), Astro (this site), Tailwind (design primitives), Plausible (analytics), Stripe (billing), Crisp (inbound chat for our own site), Loom (async demo), Descript (podcast + transcripts), Perplexity (research starting point), Ahrefs (still the best backlink index for SEO due diligence).

IMG / PLACEHOLDER

Monthly stack cost breakdown

Actual monthly spend across the 27 tools, sorted by cost and payback multiple. Swap with the live spreadsheet view.

The three we cancelled this quarter

Any honest stack report names the cancels. We killed three in the last 90 days.

A "full-stack AI SDR" product we won't name. The write layer was fine. The send layer was fine. The research layer hallucinated at rates we couldn't accept, and the platform's prompt surface wasn't editable enough to fix. We replaced it with Claude + a custom research prompt in about 4 hours.

A vector DB. We onboarded one before we had a retrieval problem. Three months later we realized every query we ran was better served by a Postgres full-text search against a few thousand rows. The vector DB solved a problem we didn't have.

A "category leader" prospecting tool. Great data, but the pricing outpaced our use case and the API access was gated behind a sales cycle we didn't want to run every renewal. Replaced with Apollo + custom enrichment for a tenth of the cost.

What we don't pay for

Three categories we've resisted so far, on purpose.

Agent frameworks. LangChain, CrewAI, AutoGen. We've tried them all and keep ending up writing thin custom orchestrators because the abstractions leak when you try to debug an actual production run. Temporal + Python is faster.

All-in-one marketing platforms. Every suite that promises "AI content + outbound + CRM + analytics" sacrifices depth in every layer to stay bundled. We prefer four best-in-class tools stitched by code we own.

Paid communities. Not because they aren't useful; because the leverage curve on tools is steeper per dollar for the way we work.

How to read this list

Do not buy these twenty-seven tools tomorrow. Start with Claude, Postgres, and Resend. That stack can ship the first revenue agent, the first dashboard, and the first send pipeline. Add a tool only when an ongoing pain point explicitly justifies it — not because you read a stack report on a Tuesday.

A working stack is usually a small stack that the team understands deeply. A dysfunctional stack is usually a large one that the team is afraid to prune.

The tool list is a footnote. The system built on top of it is the asset. If you cut every tool on this page and kept only the five in Tier A, you could still rebuild everything we've shipped. The tools don't make the system; the system makes the tools worth paying for.