AI/Nov 11 edition/Podcast
AgentsInfrastructureBusinessCoding

AgentsAutonomy & tool use

Resend CEO Zeno Rocha says products should be API-first for AI agents as the company's MCP server reaches about 1.2 million calls in August

Resend CEO Zeno Rocha argues developer products should be API-first so agents, CLIs and people can each use them their own way. He cites rapid growth in his company's MCP server and its work to make web content machine-readable.

From the Latent Space podcast episode of November 10, 2026 and cited sources: Rocha calls the right design 'formless': the full product is exposed through an API so any client can drive it. [1] [2] [4] [5]

SOURCES[1] Designing for Agents — Zeno Rocha, Resend lspace.beehiiv.com[2] Transcript: Designing for Agents — Zeno Rocha, Resend rss.beehiiv.com[4] Cloudflare: Code Mode MCP blog.cloudflare.com

AI-assisted, not human-edited. Methodology Corrections Who owns this site

KEY FACTS
What happened
Rocha calls the right design 'formless': the full product is exposed through an API so any client can drive it. [1] [2] [4] [5]
Original source
Latent Space podcast episode, independent perspective
Brief by
AT Pulse research desk AI-assisted; publisher and contact on the desk page
When
November 10, 2026 podcast episode published
Evidence
11 sources across 7 sites; transcript sourced

Reach: not measured yet

Illustration from Latent Space: Resend CEO Zeno Rocha says products should be API-first for AI agents as the company's MCP server reaches about 1.2 million calls in August
Image: Latent Space — Original article ↗
THE CORE IDEAS4 TAKEAWAYS
01

Rocha calls the right design 'formless': the full product is exposed through an API so any client can drive it. He says Resend's MCP server roughly doubled each month and reached about 1.2 million calls in August. These usage figures are company-reported. The server now has about 80 tools, while Cloudflare covers its whole API with two. Rocha says large tool lists rarely confuse models now. He worries more about tool calls returning too much data. [1] [2] [4] [5]

02

Covering the last share of what a dashboard does through the API and MCP is the hardest part, according to Rocha. That includes billing, usage, team invites, account provisioning and deletion. The interview also faults web frameworks that tie data directly to pages, because teams then have to build and maintain a public API separately. [1] [2]

03

To make pages readable by LLMs, Resend adds FAQs to most pages, publishes llms.txt and Markdown versions of pages, and supports content negotiation. It also repeats JavaScript-widget data as plain text. Rocha says an AI tool once misstated Resend's pricing because the numbers lived only in a widget. [1] [2]

04

Rocha describes deliverability as a scheduling and abuse-control problem. Resend puts the most engaged recipients first and spaces batches across mailbox providers. It ramps up new senders gradually and balances vetted IP pools. Tracked links and images use each sender's own domain to avoid shared-domain blacklisting. New senders are trusted by default but blocked quickly when bounce or complaint signals rise. [1] [2]

WHY IT MATTERS

Rocha says more work now starts in agent clients such as Claude Desktop and Codex. Implication: if that trend holds, full API and MCP coverage and LLM-readable pages could shape which developer tools agents pick and recommend.

Executive brief

Resend is a developer email-infrastructure company. In a Latent Space interview, CEO Zeno Rocha says its MCP server (the interface AI agents use to call its API) grew about 2× month over month and reached roughly 1.2 million calls in August. Cloudflare has taken the opposite approach and covers its whole API with two (episode, ~00:38–00:45).

Read the full section

Resend is a developer email-infrastructure company. In a Latent Space interview, CEO Zeno Rocha says its MCP server (the interface AI agents use to call its API) grew about 2× month over month and reached roughly 1.2 million calls in August. He says that server now exposes about 80 tools. Cloudflare has taken the opposite approach and covers its whole API with two (episode, ~00:38–00:45). Rocha also gives company figures: 4.3M users, 120,000+ paid customers and about $10M ARR. These are all self-reported and could not be corroborated. His main argument is that products should be "formless," meaning API-first so that agents, CLIs and humans can each use them in their own way. He also argues that marketing pages should be readable by LLMs.

What changed and event timeline

  1. Series A

    Resend raised an $18M Series A. Its handbook lists this after a $3M seed in April 2023 and 1,000 paying customers in December 2023 ().

  2. Cloudflare's two-tool MCP

    Cloudflare's Code Mode exposes more than 2,500 API endpoints through just search() and execute(). The tool definitions take about 1,000 tokens, versus about 1.17M tokens for one tool per endpoint ().

  3. Official Resend MCP server

    The server covers 10 tool groups, including emails, contacts, broadcasts, domains, webhooks and received mail. It supports local stdio and remote HTTP transports ().

  4. 3M users

    Resend reported 3M users, 43 staff and weekly npm downloads up from 1.4M to 6.8M. It said MCP coverage of its API rose from about 8% to about 90% ().

  5. Inboxes and Forge

    At its Resend Forward conference, the company launched Inboxes, which are shared inboxes for agents and humans, and Forge, its own mail transfer agent running on company-owned IP blocks (;).

  6. Episode published

    The interview appears to have been recorded before the October launches, because Rocha describes the infrastructure and inbox products as still "soon" (, 00:00:49, ~00:53:37).

Capabilities and access

  • It handles sending, contacts, broadcasts, domains with DNS verification, webhooks and inbound email (changelog).
  • CLI: 53 commands (Resend blog). Rocha says job applicants can apply through the CLI, the API, or Claude using the CLI's built-in skills (~00:46:46).
  • Inboxes: available in the dashboard, API, CLI and MCP (Resend on X).
Read the full section
  • MCP server: local and remote transports. It handles sending, contacts, broadcasts, domains with DNS verification, webhooks and inbound email (changelog).
  • CLI: 53 commands (Resend blog). Rocha says job applicants can apply through the CLI, the API, or Claude using the CLI's built-in skills (~00:46:46).
  • Inboxes: available in the dashboard, API, CLI and MCP (Resend on X). A competitor's comparison says the built-in agent only drafts replies, while external agents can send (Drag).
  • Forge: generally available, according to the company.

Technical analysis for researchers and developers

  • Deliverability is a scheduling problem. Rocha describes putting the most engaged recipients at the front of the queue and spacing batches across mailbox providers.
  • Rewriting tracked links to a domain shared by many customers creates a "noisy neighbor" risk: one abuser can get the domain blacklisted for everyone.
  • Rocha says he no longer sees models get confused by large tool lists. Cloudflare's design limits context cost by construction instead (Cloudflare).
Read the full section
  • Deliverability is a scheduling problem. Rocha describes putting the most engaged recipients at the front of the queue and spacing batches across mailbox providers. Ramp-up is throttled for new senders, and sending is balanced across IP pools whose history is vetted (~00:08:53–00:10:41).
  • Link domains. Rewriting tracked links to a domain shared by many customers creates a "noisy neighbor" risk: one abuser can get the domain blacklisted for everyone. Resend moved to links and images that match each sender's own domain (~00:14:23).
  • Tool count vs. context size. Rocha says he no longer sees models get confused by large tool lists. He worries more about tool calls returning too much data for the client (~00:42:56). Cloudflare's design limits context cost by construction instead (Cloudflare). No public evaluation compares the two approaches.
  • Sending path rewritten. The sending path moved from Node/TypeScript to Go to improve performance (~00:37:01).
  • Domain setup. Resend uses Domain Connect, an IETF draft for OAuth-like, one-click DNS provisioning (IETF draft).

Claims and evidence

All company figures above come from Resend. No independent audit, third-party usage data or benchmark corroborates them. Rocha himself notes that SendGrid still sends far more email (~00:28:40).

Read the full section
ClaimEvidence
4.3M users; 120k+ paid customers; ~$10M ARRInterview only, ~00:03:32 and 00:00:26
3M users in June 2026Resend blog
MCP reached ~1.2M calls in August; first version built by Yoko Li of a16z, an investorInterview, ~00:38:46–00:39:39
Official server credited to Derich Pacheco and Felipe FreitagChangelog
Resend npm downloads now more than double SendGrid'sInterview, 00:00:05

All company figures above come from Resend. No independent audit, third-party usage data or benchmark corroborates them. Rocha himself notes that SendGrid still sends far more email (~00:28:40).

Context and prior work

  • Resend started from React Email, an open-source project launched in December 2022.
  • Incumbents include SendGrid (owned by Twilio), Postmark, Mailgun and Amazon SES. Resend itself started on SES.
  • Agent-inbox competitors include AgentMail, Nylas and Cloudflare.
Read the full section
  • Resend started from React Email, an open-source project launched in December 2022. Rocha calls it "our Next.js": a free tool that fixes email templating and leads developers to Resend's sending platform (~00:24:06; handbook).
  • Incumbents include SendGrid (owned by Twilio), Postmark, Mailgun and Amazon SES. Resend itself started on SES.
  • Agent-inbox competitors include AgentMail, Nylas and Cloudflare. According to Drag's comparison, those offer dedicated addresses for agents, while Resend and Drag use shared team inboxes (Drag).

Limitations, safety and contested findings

  • Resend trusts new senders by default, which invites abuse. Rocha cites a group that brought about 20,000 domains.
  • Drag's comparison is written by a competitor, and the author did not test the developer inboxes hands-on (Drag).
  • Rocha's claim that Claude Opus 4.5 favors newer tools ("recency bias") is his own observation, not a measured result (~01:06:26).
Read the full section
  • Abuse risk. Resend trusts new senders by default, which invites abuse. Rocha cites a group that brought about 20,000 domains. Resend relies on fast blocking based on bounce and complaint signals (~00:11:03, ~00:16:22).
  • Weak comparison evidence. Drag's comparison is written by a competitor, and the author did not test the developer inboxes hands-on (Drag).
  • Anecdotal model behavior. Rocha's claim that Claude Opus 4.5 favors newer tools ("recency bias") is his own observation, not a measured result (~01:06:26).

Business and practitioner implications

  • Aim for full coverage of what the dashboard can do through the API and MCP.
  • Make content machine-readable. An AI tool got Resend's pricing wrong for that reason (~00:59:35–01:04:48).
  • Per-contact pricing breaks down for fast-growing AI customers such as Mistral, according to Rocha (~01:15:21).
Read the full section
  • Design for agent users. Aim for full coverage of what the dashboard can do through the API and MCP. Rocha says the last ~10% is the hardest: billing, team invites, provisioning and deletion (~00:43:47). Avoid frameworks that tie data to pages so tightly that no public API exists.
  • Make content machine-readable. Rocha's practices include FAQs on every page, llms.txt, .md versions of pages and content negotiation. Data shown only in JavaScript widgets should also appear as text, because agents often don't run JavaScript. An AI tool got Resend's pricing wrong for that reason (~00:59:35–01:04:48).
  • Pricing pressure. Per-contact pricing breaks down for fast-growing AI customers such as Mistral, according to Rocha (~01:15:21).
FOLLOW THE EVIDENCE

The source trail.

Sources (11)
A LITTLE LESS NOISE. A LOT MORE CONTEXT.

Stay curious.
Follow the evidence.

Independent perspectives, the original sources, and room for the questions that don't have easy answers.

How we build the brief
Connect with us

Find us where you already read.