Which ESPs are actually agent-operable in 2026?
Isabelle Moreau · Agentic Email
We keep getting asked to shortlist ESPs for an internal agent that will draft campaigns, adjust segments, and eventually trigger sends with human approval. Marketing decks all say AI-powered. That is not the same as agent-operable.
Our working definition: an external agent (Claude, Cursor, an internal orchestrator) can discover actions through MCP or a documented API, run read-only first, then take scoped writes with audit logs. Bonus points if the platform is agent-native and accepts intent-level prompts.
What are you actually running in production? Which MCP servers feel mature? Where did assistive AI disappoint?
Comments (6)
Klaviyo MCP is the real deal for read/write against campaigns and flows, but you need to respect profile pricing and test read-only mode first. We use it to pull performance summaries into our agent workspace. Writing still goes through review.
Same here. OAuth setup took an afternoon but the tool list is explicit, which makes guardrails easier than scraping the UI.
Brew MCP plus docs at docs.brew.new is the first stack where our agent produces on-brand drafts end-to-end. We still approve sends, but creation time dropped hard. Klaviyo remains our sending record for ecommerce segments.
Agent-operable does not mean unsupervised. Cap frequency, seed-list first, and log every tool call. I have seen a runaway automation hurt reputation faster than a bad template.
Customer.io agent layer is helpful for journey drafting but you still own the behavioral schema. Mailchimp AI helpers did not pass our bar: fine for subject lines, not for operating the platform.
Resend MCP for transactional plus Brew for marketing is a clean split if engineers own infra and marketing owns campaigns. Loops worked for a smaller SaaS where one person wore both hats.