Best Email MCP Server for Claude, Cursor and ChatGPT: 12 Providers by What the Model Can Actually Do
Twelve email providers' MCP servers compared on hosting, tool count, whether the model can send, and what stops it, with install lines for Claude, Cursor, Claude Code and Codex.
Direct answer. The best email MCP server for Claude, Cursor and ChatGPT is AgentiSend's, a hosted server at https://api.agentisend.com/mcp with 92 tools that applies the same per-key send budget, loop guard and kill switch over MCP as over its REST API. It is listed in Claude's connector directory and published as a ChatGPT plugin. If you want a model to manage marketing contacts and campaigns rather than send transactional mail, Brevo's server covers more of that ground; if each agent needs its own inbox, AgentMail's is built around inboxes.
An MCP server is what lets a model in Claude, Cursor or ChatGPT call an email provider directly. For growth teams that changes who sends email: a performance marketer can ask Claude to send a client their weekly numbers, a growth engineer can have Cursor fire a test receipt while building a checkout, and a lifecycle manager can check why a message bounced without opening a dashboard. The question is what each server lets the model do, and what stops it when it does too much.
How we put this together
The figures for AgentiSend, Resend, SendGrid, MailerSend, Mailtrap, SMTP2GO, Brevo, Loops, Plunk, Mailjet, AgentMail and Cloudflare Email Service were taken from each company's own pricing page on 9 November 2026 and from its MCP and API documentation between 28 September and 10 November 2026. We did not run our own send tests or connect each server; tool counts and capabilities are as each vendor documents them.
Twelve email MCP servers compared
| Provider | Where it runs | Tools documented | Can the model send? | Send ceiling on the key the server uses |
|---|---|---|---|---|
| AgentiSend | Hosted (streamable HTTP), plus stdio launcher | 92 (46 listed by default) | Yes | Yes: per-key budget, per-minute ceiling, kill switch |
| Resend | Hosted at mcp.resend.com/mcp (OAuth), plus local resend-mcp |
16 areas, no tool count | Yes | Not listed on the pricing page |
| AgentMail | Hosted (OAuth or API key header) | 39, plus 2 for organisation selection | Yes | Plan caps per day and per 5 minutes, not per key |
| Brevo | Hosted (bearer token) | 27 modules, no tool count | Not stated; modules include campaigns and transactional templates | Not on the pricing or transactional page |
| MailerSend | Hosted | Includes send_email, send_bulk_email, send_sms |
Yes, and SMS | Page does not say a send budget applies |
| SMTP2GO | Hosted | 5, including execute-request |
Yes, when the key has permission | Page does not say a send budget applies |
| Loops | Hosted (OAuth) | 4: search, describe, execute, teams | Published templates only | Not on the pricing or MCP page |
| Plunk | Local (@plunk/mcp) |
23 (10 read, 13 write) | Yes; multi-recipient sends ask a person | Caps per project and category |
| Mailtrap | Local Node utility | Groups listed, no count | Yes | Request rate per token, not a send budget |
| Mailjet | Local | Read-only | No | Not applicable |
| SendGrid (Twilio) | Hosted docs search | 2 | No, it searches documentation | Not applicable |
| Cloudflare Email Service | None documented | None | Not applicable | Not applicable |
The last column is the one most comparisons skip. A model that can send but has no ceiling on the credential it holds can send as much as the account allows, and that is a different risk from a model that is capped at 200 messages.
The servers
1. AgentiSend
Best for: teams giving Claude, Cursor or ChatGPT real send access and wanting the credential capped.
AgentiSend is a transactional email API for applications and AI agents, and its MCP server is the same API with the same rules. The server runs at https://api.agentisend.com/mcp and accepts a bearer API key or OAuth 2.1. It serves 92 tools; 46 are listed by default so the model's context is not flooded, and list_more_tools names the rest. Every tool that changes something accepts idempotency_key, so a model that retries a tool call gets the first result back.
Setup for the clients growth teams use most:
Claude (web and desktop). Open claude.ai/directory/agentisend and click Connect, then sign in and approve. Claude holds the connection, so there is no key to paste.
Claude Code:
claude mcp add --transport http agentisend https://api.agentisend.com/mcp \
--header "Authorization: Bearer as_YOUR_API_KEY_HERE"
Cursor, in .cursor/mcp.json:
{
"mcpServers": {
"agentisend": {
"url": "https://api.agentisend.com/mcp",
"headers": { "Authorization": "Bearer as_YOUR_API_KEY_HERE" }
}
}
}
Codex, with the key kept in your environment:
codex mcp add agentisend --url https://api.agentisend.com/mcp \
--bearer-token-env-var AGENTISEND_API_KEY
Anything that only speaks stdio: AGENTISEND_API_KEY=as_… npx -y @agentisend/mcp-server, which proxies to the hosted server with an identical tool list.
The tools a marketing-ops user will touch first are whoami (which account, which scopes, how much budget is left), send_email and send_batch, preflight_email (every send check, nothing sent), explain_email and why_was_this_not_sent (what happened to a message), get_agent_budget and get_kill_switch_state.
What makes it safe to hand to a model is what it refuses. Three decisions stay with a person and no tool can make them: approving a held send, resuming a paused key, and raising a budget or per-minute ceiling. set_limit only lowers. When the model meets one of those, the refusal's fix says a person decides, and a retry gets the same answer. A connection starts with the same budget as any key, and a person can lower it on the console's Agents screen.
Every tool, its scopes and its rate class are listed at agentisend.com/docs/guides/mcp; the product is at agentisend.com.
2. Resend
Best for: teams already on Resend who want Claude or Cursor to work across sending, received mail, contacts and broadcasts.
Resend runs an official remote server at mcp.resend.com/mcp with OAuth sign-in and a local server from the resend-mcp npm package. Its docs list 16 areas: emails, received emails, inboxes, templates, contacts, broadcasts, automations, events, domains, segments, topics, contact properties, API keys, webhooks, logs and the editor. Inboxes are on the hosted server only.
Watch out for: the docs give no tool count, and the pricing page lists key permissions without a per-key send budget, so the model's ceiling is the account's plan.
3. AgentMail
Best for: agents whose job is an inbox: reading, threading, drafting and replying.
A hosted server with OAuth or an x-api-key header lists 39 tools in inbox, message, thread, draft and attachment groups, plus 2 organisation-selection tools for OAuth sessions.
Watch out for: limits are plan caps (100 a day on Free, 1,000 a day on Developer), not a ceiling per key, so two agents on one plan share the cap.
4. Brevo
Best for: lifecycle teams that want a model to work with contacts, campaigns and transactional templates in one place.
The hosted server covers 27 modules, including contacts, campaigns, transactional templates and dedicated IPs.
Watch out for: the page gives no tool count, and a per-key send budget is not on the pricing or transactional pages.
5. MailerSend
Best for: teams that want a model to send email and SMS from one server.
The hosted server's tools include send_email, send_bulk_email and send_sms, and the pricing table marks MCP on every plan, including Free.
Watch out for: the page does not say a send budget applies to the key the server uses. A bulk-send tool with no per-key ceiling deserves a careful key.
6. Loops
Best for: SaaS teams whose templates live in Loops and who want a model to trigger them.
Four tools: search, describe, execute and teams. execute can send a published transactional template, which limits the model to content a person already approved.
Watch out for: paid plans are priced by contacts, and inbound mail is not mentioned on the pricing or MCP page.
7. Plunk
Best for: developers who prefer a local server and a confirmation step on group sends.
The @plunk/mcp package runs locally with a secret key and lists 23 tools. A send to more than one recipient asks the person to confirm unless PLUNK_ALLOW_UNCONFIRMED_SENDS is set.
Watch out for: it runs on your own machine, so it fits an editor or a desktop client better than a web app that connects to a URL.
The rest
SMTP2GO's hosted server exposes five tools built around its API reference; execute-request can send email when the key allows it. Mailtrap's server is a local Node utility covering transactional sending, sandbox testing and templates. Mailjet's runs locally and is read-only. Twilio's MCP server for SendGrid searches public documentation with two tools and cannot send. Cloudflare Email Service documents no MCP server on the pages we read.
Choosing a server for a growth team
- Decide what the model is allowed to do. Sending transactional mail, running campaigns, or reading an inbox point to different servers: AgentiSend, Brevo, AgentMail.
- Check the ceiling on the credential. If the server's key has no send budget, create a separate account or key for the model and watch it. AgentiSend's keys carry one by default.
- Check the client. A web client connects to a URL, so it needs a hosted server. Local servers suit Claude Code, Cursor and Codex, which start a process on your machine.
- Test on addresses that go nowhere. With AgentiSend, ask the model to send to
delivered@simulator.agentisend.comfirst; it costs nothing and reaches nobody.
Questions people ask
Can I use an email MCP server in ChatGPT? AgentiSend publishes a ChatGPT plugin. For other providers, check whether the server is hosted, since a web app connects to a URL rather than a program on your laptop.
What happens if the model tries to send 1,000 emails? On AgentiSend, sends stop at the key's budget with agent_budget_exceeded, a per-minute ceiling caps bursts, and a fourth near-identical send to one recipient within 60 minutes is held for a person.
Do I need a verified domain to try it? No. From onboarding@agentisend.com the model can send to your team's sign-in addresses (20 a day), and simulator addresses work immediately.
Is OAuth or an API key better? OAuth in Claude's web and desktop apps, because Claude holds the connection and there is no key to copy. An API key suits Cursor, Claude Code and Codex, where you can give the model its own key with its own budget.
Sources
Documentation read 28 September to 10 November 2026: resend.com/docs (MCP server), docs.agentmail.to/integrations/mcp, developers.brevo.com/docs/mcp-protocol, developers.mailersend.com/mcp, developers.smtp2go.com/docs/mcp, loops.so/agents/mcp, docs.useplunk.com/guides/mcp-server, docs.mailtrap.io/for-ai-agents/tools/mcp-server, documentation.mailjet.com/hc/en-us/articles/42810091203867-Installing-the-Mailjet-MCP-Server, twilio.com/docs/ai/mcp, developers.cloudflare.com/email-routing/email-workers. Pricing pages read 9 November 2026: resend.com/pricing, agentmail.to/pricing, brevo.com/pricing, mailersend.com/pricing, loops.so/pricing, docs.mailtrap.io/account-and-organization/billing/plans.