Claude Code chose Resend in 64 out of 102 cases: how to build products for agents

Imagine asking Claude Code to “add email sending,” accepting the patch, and finding a specific provider already integrated into your repository. You didn’t choose it. The agent did. That is the moment Zeno Rocha, co-founder and CEO of Resend, examines on the Latent Space podcast. The discussion began with a question puzzling the hosts: why do agents recommend Resend instead of Cloudflare Email, despite Cloudflare’s home-field advantage and scale?

First in the recommendation wins

In one AEO study (answer engine optimization), Claude Code named Resend first in 64 out of 102 extracted recommendations. SendGrid came first just seven times, but appeared 55 times as an alternative. The logic is straightforward: once a developer accepts the patch, the second option rarely gets a look.

An important caveat: a newer, broader study showed a more competitive field, with Postmark leading across most models. So the ranking is not permanent and depends on the model and its version. At the time of recording, Rocha said Resend had 4.3 million users and 120,000 paying customers.

A product an agent can read

Resend’s first distribution channel was not its API but React Email, a library for building email layouts. It also works with SendGrid, and the team even documented how to use them together. By the time a developer needed a different provider, some of Resend’s code was already in their repository.

Now look at it through an agent’s eyes. Resend’s pricing page has a slider: a person adjusts it to the volume they need and sees the price. That is of little use to an agent. The solution is modest: append .md to the page URL to get a plain table that can be read in a single request. A useful self-test: open your own pricing page and check whether a model can use it to answer a customer’s question.

80 tools or two

Resend’s MCP server grew from around 20 tools to more than 80. Rocha’s goal is to make everything available in the dashboard accessible through chat: send an email, invite a colleague, check usage, manage billing. Sending an email is useful for a demo; the real work lies in the long tail of these operations.

Cloudflare takes the opposite approach: Code Mode exposes its entire API through two tools, search() and execute(). But tool count is not the whole story. Rocha expands the collapsed calls in his own sessions to see what the agent actually did. In one case, the result simply did not fit into the context, and reducing the number of tools would not have fixed that.

A low-cost experiment worth repeating: expand the response from a call you make every day and see how much of it the model actually needs. In its guidance on writing tools for agents, Anthropic also recommends returning less data when the result overwhelms the context window.

After the choice comes delivery

Persuading an agent to use your SDK is half the battle; you still have to deliver the email. Click tracking is one example. A provider may rewrite links through a shared domain to count clicks. If one customer uses that domain for phishing and it gets blocklisted, legitimate emails from other customers will suffer too. That is why Resend introduced custom tracking domains.

Rocha also explains why he wants to retain the free plan, even though spammers may sign up for it. The company addresses the risk through anti-abuse processes and a stated $3 million third-party security guarantee.

What to do with this

The limitations are clear. AEO studies measure specific models at a particular moment, and the podcast is a founder’s account, not an independent audit. Resend’s results cannot simply be applied to every product.

But the practical takeaways hold up:

1. Create Markdown versions of your documentation and pricing page so agents can read them without interactive elements.
2. Check your MCP server’s responses for bloat: the problem is often the size of the result rather than the number of tools.
3. Give developers a useful artifact before they choose a provider, as React Email did.
4. Avoid shared resources where one bad customer can damage everyone’s reputation.

The central point: a model is now often your first user, and your product needs to make its decision straightforward and sound.