Claude Code Picks Resend First for Email in 63% of Test Cases, SendGrid in 7%

TL;DR

When a developer asks an agent to wire up email, the agent usually names Resend before any human weighs in. In Amplifying's study of 2,430 open-ended prompts to Claude Code, Resend was the first recommendation in 64 of 102 extractable email answers, 62.7%, against 7 first picks for SendGrid, which still appeared as the suggested alternative 55 times. On the Latent Space podcast (Nov 10), Resend co-founder and CEO Zeno Rocha walked through the company side of that number: 4.3 million users, 120,000 paying customers, and a distribution strategy built on making the product readable by machines. It is the most detailed public account so far of agents working as an acquisition channel.

What the study measured

Amplifying ran 2,430 prompts across four greenfield repos, a Next.js SaaS, a FastAPI backend, a React SPA and a Node CLI, with no tool name in any prompt. Just openers like "i need a database, what should i use." Three Anthropic models, Sonnet 4.5, Opus 4.5 and Opus 4.6, ran three times each over a clean git checkout between every prompt, and 2,073 responses produced an identifiable primary tool. The email category yielded 102.

Across all 20 categories the loudest finding was that agents build rather than buy. Custom implementations took 252 of 2,073 primary picks, 12%, more than any single vendor collected. Email is one of the categories where a vendor still wins anyway. Resend's pick rate ran between 66% and 84% depending on the model, and the authors flag their own caveat: they cannot separate product quality from training-data frequency. A tool recommended nine times out of ten may be better, or it may simply show up more often in what the model once read.

The playbook, in the order Resend built it

React Email came first. Before the sending API existed, the team shipped an open-source template library and documented how to send through those templates with SendGrid. Developers adopted the templates, so by the time they wanted a different provider, Resend code already lived in the repo. The wedge predates the product it markets.

Then the surfaces went machine-readable. The pricing page has a volume slider for humans; append .md to the URL and the agent gets a table it can read in one request. I checked: resend.com/pricing.md returns the full plan grid as markdown, free tier at 3,000 emails a month (100 a day) against Pro at $20 for 50,000. The docs serve the same format, with a llms.txt index at the root for crawlers that want the map before the territory.

The third piece is the MCP server: a hosted endpoint at mcp.resend.com, OAuth on connect, plus plugins that install it into Claude, Claude Code, Cursor, Codex and more than a dozen other clients. It has grown from roughly 20 tools at launch to over 80, because Rocha wants someone in a chat client to do everything the dashboard does: send, invite a teammate, read usage, manage billing.

80 tools, or two

That tool count puts Resend at one end of a live argument about MCP design. Cloudflare's Code Mode server sits at the other: two tools, search() and execute(), covering an API of 2,500+ endpoints in roughly 1,000 tokens of context. Cloudflare claims a 99.9% cut against the 1.17 million tokens an equivalent per-endpoint server would load. The model writes JavaScript against a typed spec inside a sandboxed Worker instead of choosing among hundreds of predefined tools.

Rocha's counter-evidence is an experiment anyone can repeat. He opened the collapsed tool calls in his own sessions and found one result too large to fit in the context window. Cutting the tool count would not have shrunk that response, so Resend trims what tools return instead of how many exist.

The pick is only half the job

Getting installed means little if the mail lands in spam. Click and open tracking normally run through a provider's shared domain, so one customer's phishing can poison another customer's links. Custom tracking domains, free on every plan since April, move the redirect onto a domain the sender owns. Rocha also keeps the free tier open although spammers can sign up, and the company has put $3 million into making email safer.

Enterprise discovery runs on the same fuel. He found Warner Bros. Discovery in the customer database on a Saturday by sorting for short .com domains, which may be the cheapest enterprise lead-gen exercise ever recorded.

The catches

• The study is a February snapshot of three Anthropic models and one assistant. Opus 4.6 already picked Resend less often than the 4.5 models did (66% against an 81% 4.5-generation average), and these numbers move with every release.
• Latent Space's own broader tracker, cited on the podcast, is more competitive: Postmark finishes first for most models there. Resend's lead is real but not universal.
• Extraction was LLM-based, subagents reading Claude Code's own responses, a bias the authors call out themselves.
• The 4.3 million users and 120,000 paying customers are Rocha's figures from the recording. No audited filing sits behind them.
• A high pick rate can track training-data frequency as much as quality; the study's authors say exactly that.

Why builders should care

Distribution is moving to a channel where you cannot buy ads. A developer asking an agent for an email provider is like a tourist asking a local where to eat: the billboard stops mattering the moment the person you asked has an opinion. And the study's stickiness is mechanical, not psychological. If the developer accepts the patch, the second choice never gets tried. SendGrid's 55 appearances as the alternative made it the agent's "also fine," and nobody's quarter is built on also fine.

The checklist transfers to any product with an API. Mirror your pricing and docs as markdown, publish a llms.txt index, ship an MCP server plus a plugin for the big clients, then measure the way the study did: open-ended prompts, no tool names, several runs. Your marketing site now serves two audiences, and one of them never sees the hero image.

Key Takeaways

• In Amplifying's study of 2,430 open-ended Claude Code prompts, Resend was the first email recommendation in 64 of 102 extractable answers (62.7%, 95% CI 53.1-71.5). SendGrid led 7 times (6.9%) while appearing as the alternative 55 times.
• Resend's CEO says the company has 4.3 million users and 120,000 paying customers, and traces agent preference to machine-readable surfaces (pricing.md, llms.txt, markdown docs), the React Email open-source wedge, and an 80-tool MCP server with plugins for the major agent clients.
• The study's biggest cross-category finding: agents build rather than buy. Custom implementations took 252 of 2,073 primary picks, 12%, more than any single tool. Email is a category where a vendor still wins.
• MCP design is contested: Resend's 80+ tools aim for chat parity with the dashboard; Cloudflare's Code Mode covers 2,500+ endpoints with two tools in about 1,000 tokens. Rocha's field data says response size, not tool count, drives context cost.
• Caveats: one assistant, three models, a February snapshot; Postmark tops Latent Space's broader tracker for most models; user and customer figures are unaudited CEO claims; pick rates may reflect training data as much as quality.
• The work after the pick matters as much as the pick: custom tracking domains, free on every plan since April, keep one customer's abuse from damaging another customer's deliverability.

Sources: Designing for Agents, Zeno Rocha, Resend, Latent Space (Nov 10, 2026), What Claude Code Actually Chooses, Amplifying (Feb 2026), Resend pricing, markdown mirror, Resend MCP Server docs, Resend llms.txt, Code Mode: give agents an entire API in 1,000 tokens, Cloudflare blog, Custom Tracking Domains, Resend (Apr 16, 2026), Resend's $3M email safety commitment, Latent Space Frontier AEO tracker