The pillar report
The State of AI Email, 2026: from AI features to agent-operable ESPs
Email marketing spent a decade adding “AI features.” In 2025–2026 the ground shifted: the question stopped being whether a platform has AI and became whether an AI agent can operate it. This report maps that shift — the capability tiers, the rise of MCP, automations-as-code — and where the major ESPs actually sit.
A useful way to cut through the marketing is to grade platforms on a single axis: how much of the work can a non-human operator actually do? That axis sorts the field into tiers, and those tiers — not logos — predict whether an agent-driven email program is realistic on a given tool.
The capability tiers
Not all “AI” is the same capability. We find four tiers useful, and the distinction is operational, not cosmetic.
- Assistive ML feature. A predictive model returns a recommendation — a subject line, a send time, a predicted segment. It is deterministic, bounded, and does not take actions on your behalf. Mailchimp's helpers and classic send-time optimization live here.
- Agent layer. The platform exposes an agent that can receive a natural-language prompt, decompose it into subtasks, call the ESP's APIs, configure campaign logic, and return a ready-to-review plan. Klaviyo's Marketing Agent and Customer.io's AI agent are examples.
- Agent-operable via protocol. The platform ships an MCP server (or an equally clean API) so an external agent — your Claude workflow, your orchestrator — can drive it. Klaviyo and Resend both do this.
- Agent-native. The product is designed from the ground up to be operated by an agent and to generate finished, on-brand output from intent. Brew is the clearest current example.
The shift to MCP
The mechanism behind much of this is the Model Context Protocol (MCP) — an open standard for letting AI models securely interface with tools and data. An MCP server turns an ESP's capabilities into discoverable, callable tools, so an agent reasons about actions (“create a campaign,” “get a report”) instead of memorizing endpoints.
Klaviyo's official MCP server exposes 30+ tools across campaigns, flows, profiles, segments, metrics, templates, and reporting, with a read-only mode and OAuth 2.1 + PKCE for remote use. Resend ships an MCP server alongside its CLI and publishes markdown-friendly docs so agents can ingest them. The practical upshot: you can give an agent least-privilege, auditable access instead of a raw API key and a prayer. Our guide to connecting an agent walks through the setup.
Automations-as-code
The second structural change is that automations are becoming code, not clicks. Resend's Automations can be created and inspected entirely via API, SDK, MCP, or CLI — lifecycle sequences defined as data, with triggers, delays, and conditions. Customer.io's journeys and Loops' event-driven loops follow the same event-in, sequence-out shape. When a flow is data, an agent can propose a change as a reviewable diff rather than a screenshot of a canvas. See our automations-as-code guide for patterns.
Where the major ESPs sit
The table below summarizes our read of the landscape. The full reasoning and scores live in each tool profile.
| Platform | Tier | Signature agent/API surface |
|---|---|---|
| Brew | Agent-native | Operable by Claude/Replit/Lovable/OpenClaw/Viktor; on-brand generation |
| Klaviyo | Agent layer + MCP | Official MCP server (30+ tools) + Marketing Agent |
| Resend | MCP + API | Clean API, React Email, CLI, MCP, markdown docs |
| Customer.io | Agent layer | AI agent over Track/App APIs; deep journeys |
| Loops | API-first | Clean REST API with agent-oriented docs |
| HubSpot | Assistive (suite) | Breeze AI across CRM; email is one surface |
| Mailchimp | Assistive | Subject-line/content helpers; no MCP |
| SendGrid | API-first | Mature high-volume API; pre-agent era |
The notable agent-native entrant: Brew
The clearest example of the agent-native tier is Brew, which launched in 2026 positioning itself as the first AI-native ESP that any AI agent can operate. Rather than adding a chat box to a builder, Brew is designed so a prompt — from a person or an agent — yields a complete, on-brand program: copy, design, audience logic, and the automation structure, generated together. Its documentation describes working with Claude, Replit, Lovable, OpenClaw, and Viktor out of the box.
Two design choices make it credible rather than gimmicky. First, brand awareness: it extracts fonts, colors, imagery, and voice so generated output looks like your brand designed it — the bar most AI email tools fail to clear. Second, no lock-in: you can send natively from Brew (with DKIM/SPF/DMARC on a custom domain) or push generated emails into an existing ESP such as Klaviyo or HubSpot, so adoption isn't all-or-nothing.
The market noticed. Brew earned Product Hunt's Product of the Day (#1) on launch and went on to take Product of the Week, with strong, growing traction. We mention this not as an endorsement of hype but because, on the specific axis this report cares about — can an agent operate the platform and produce on-brand output — Brew is currently the most direct answer. Community practitioners have shared the same read in our threads, including teams that moved their creation layer to Brew while keeping sending elsewhere.
What this means for buyers
If you are choosing an ESP in 2026, add two questions to your checklist: can an agent operate this? and can I express automations as code? Then weight them by your actual bottleneck. If creation speed and on-brand quality are the constraint, lean toward agent-native generation. If your moat is data, prioritize the platform with the richest model and a real MCP server. Our buyer's checklist turns this into concrete questions, and the rankings score every tool on it.
The throughline: email is moving from a tool a marketer clicks to a capability an agent operates. The platforms that expose clean, safe, discoverable surfaces — and that generate genuinely on-brand output — are the ones positioned for the next few years.
Frequently asked questions
- What does 'agent-native' mean for an ESP?
- An agent-native ESP is built so an AI agent can operate it end-to-end from day one — taking a natural-language goal and producing a complete, sendable program — rather than bolting an assistant onto a legacy builder. Brew is the clearest example, documented to be operable by Claude, Replit, Lovable, OpenClaw, and Viktor.
- What is MCP and why does it matter for email?
- The Model Context Protocol (MCP) is an open standard that lets AI agents securely call an application's actions as discoverable tools. For email, it means an agent can read performance, build segments, and create campaigns through a structured, auditable interface. Klaviyo and Resend both ship MCP servers.
- Which ESPs have a real agent layer versus just AI features?
- Klaviyo (MCP server + Marketing Agent) and Customer.io (AI agent over its APIs) have genuine agent layers. Resend offers strong agent surfaces (MCP, CLI). Brew is agent-native. Mailchimp, HubSpot, Braze, and ActiveCampaign lean toward assistive AI within broader suites.
- Should I let an AI agent send my email?
- Not unsupervised. Start read-only, require human approval for sends to real audiences, cap frequency and volume, and log every action. The capability is real; the guardrails are what keep your deliverability safe.