ComparisonUpdated October 27, 2026
Superhuman vs Command Mail: two MCP servers, two permission sets
Superhuman Mail and Command Mail both offer a hosted MCP server that lets Claude, ChatGPT or Codex work in your inbox. Superhuman's server works directly on your Gmail and Outlook accounts, can send email without delay, and describes its limits mainly as tool approvals in your AI client plus a read-only mode.[1] In Command Mail, according to its documentation, the permissions you grant when connecting and an autonomy level for each mailbox also determine what the agent may do.[3][4]
If you like Superhuman as an email client and already pay for Business, you get MCP access included; if you want to limit your agent through connection permissions and mailbox levels, Command Mail is the better fit.[1][3]
Command Mail is our product. Everything about Superhuman comes from the Superhuman help center, retrieved on 2026-10-27; we have not tested the Superhuman MCP ourselves.[1][2]
As of: 2026-10-27
Contents
- What this comparison is about
- Superhuman and Command Mail at a glance
- What Superhuman does well
- Where the Superhuman MCP reaches its limits
- How Command Mail sets the limit
- An example: the morning pass with reply drafts
- When Superhuman is the better choice
- When Command Mail fits
- Who Command Mail is not for
- Switching and setup
- Further reading
- Frequently asked questions
- Sources
01What this comparison is about
Superhuman Mail is an email client for Gmail and Outlook accounts with built-in AI and team features such as shared conversations and comments.[1][2] With the Superhuman Mail MCP Server it has also become an entry point for outside agents: your Claude or ChatGPT searches, drafts and sends through Superhuman.[1] Command Mail is an inbox you forward your existing addresses to and connect your own agent to.[3]
So both products now answer the same question: how does my agent reach my email? The differences are where the mail lives, what the agent may do, and who enforces that limit. That is the subject here, not which email client looks nicer.
02Superhuman and Command Mail at a glance
| Question | Superhuman Mail | Command Mail |
|---|---|---|
| Price (retrieved 2026-10-27) | Starter $30/month or $300/year, Business $40/month or $396/year, Enterprise on request[2] | Free €0; Pro €19/month for the first 100 paying Pro accounts while the subscription runs, otherwise €29/month (yearly €190 or €290); Team €39 per person per month[3] |
| MCP server | hosted, OAuth, Business and up; not on Starter[1] | hosted, OAuth 2.1, on every plan; plus CLI and local MCP[3] |
| Mailboxes | Gmail and Outlook accounts on your own seat[1] | any address by forwarding, up to ten mailboxes on Pro[3] |
| Sending by the agent | send_draft can send immediately; confirmation depends on the AI client[1] | directly only with mail:send from a mailbox on auto; approved drafts with mail:send-approved, not from mailboxes on read[3] |
| Setting the limit | per-tool approval in the client; read-only mode possible according to the help center[1] | permissions per connection, autonomy level per mailbox[3] |
| Calendar | yes, create events and find availability[1] | not described in the documentation[4] |
| Log of MCP actions | none, according to the help center[1] | activity log for agent and rule actions; suggestions wait there for a decision[4] |
03What Superhuman does well
Superhuman's MCP server covers more than email. It offers more than 30 tools: filtered and natural-language search, drafts in your writing style, reminders, shared threads with comments, read statuses, unsubscribing from lists and calendar events.[1] If you want to handle email and calendar in one pass, for example answering a request and proposing a meeting time, you will find matching tools; the Command Mail documentation does not describe a calendar.[4]
Setup is documented in breadth. The help center walks through Claude Code, Claude Desktop, ChatGPT, Codex as a plugin with bundled skills, Cursor, Gemini CLI, Perplexity and Amazon Quick Suite.[1] In Claude Code one command plus a browser sign-in is enough. Ready-made skills such as Morning Briefing, End of Day Wrap-Up and Batch Draft Writer live in a separate library.[1]
Several accounts can be linked in one session. The agent then searches across all Gmail and Outlook accounts on your seat and switches between them to draft and send.[1] And Superhuman is a mature email client with its own AI; Business adds Auto Drafts, Ask AI, custom auto labels and HubSpot and Salesforce integration.[2]
04Where the Superhuman MCP reaches its limits
Superhuman puts the sending limit in your AI client. The send_draft tool can send email immediately and alternatively supports Smart Send and scheduled sending; the help center explicitly advises turning on confirmation for every call in the client, because otherwise the message goes out without delay.[1] Whether that confirmation appears depends on the client. In Claude, admins and users with the right permissions can set individual tools to "Always Allow"; in ChatGPT, according to the help center, only read actions can skip confirmation, while drafting, sending and deleting still require it.[1] The help center says the server can also run read-only, but the article does not explain how to set that up.
The help article does not document an autonomy level per mailbox. It describes tool approvals in the client and a default account the agent acts on unless you name another one.[1] It also does not describe a built-in suggestion stage for trash_thread or update_thread.
Superhuman names two further limits itself. There is currently no MCP-specific logging and no identifier showing whether an action came from the agent or from you.[1] And the commitment that subprocessors do not train on customer data covers Superhuman's built-in AI; on the MCP route the data goes to your AI provider, whose policies then apply.[1] The same holds for Command Mail: what your agent retrieves is processed by its provider under that provider's terms.[5] The agent cannot send attachments through the MCP yet.[1]
And the MCP server requires Business. It is not part of the $30 Starter plan, and on Enterprise it is off by default and enabled through your Customer Success Manager.[1][2]
05How Command Mail sets the limit
In Command Mail you choose which permissions the agent gets when you connect it. mail:read reads and sorts, mail:draft writes drafts and never sends, mail:send-approved sends exactly the draft revision a person approved, and mail:send allows sending without asking.[3] Every permission can be switched off, and sending without approval starts switched off. According to the documentation, connection permissions and the autonomy level determine the allowed actions; the client's approval setting comes on top as an additional brake.[4]
On top of that, every mailbox has an autonomy level: read, suggest, rules or auto.[3] The agent may send directly only from a mailbox on auto, with a unique send key per email, a daily limit and at most 20 recipients. On suggest and rules, direct archiving, spam and trash actions become suggestions, and on rules active rules can already take effect; on both levels mail only goes out through approved drafts. On read it changes no existing mail and sends nothing, not even approved drafts; it can still save drafts. Permanent deletion and emptying Trash and Spam require both read and send permission and a mailbox on auto.[3]
Command Mail treats incoming email as untrusted data and flags suspected prompt injection. An agent does not write on its own to detected automated senders, mailing lists or no-reply addresses; detection relies on message features and address patterns and can miss individual cases.[3][4] The first call of every session, capabilities, tells the agent its permissions, the level of each mailbox and its limits, so it does not have to guess.[3]
Command Mail has limits too. The documentation does not describe a calendar, read statuses or CRM integrations.[4] Mail arrives by forwarding; what the agent archives stays where it was in Gmail or Outlook. The read permission covers every mailbox in the account, so the autonomy level does not separate read access between mailboxes. The operator is FYN Labs LLC in the US; the app and database run on Cloudflare.[5]
06An example: the morning pass with reply drafts
Say your agent should go through the inbox every morning, summarize what matters and prepare replies to three requests. With Superhuman it uses list_threads, query_email_and_calendar and create_or_update_draft; the Morning Briefing skill covers the overview and the Batch Draft Writer skill the replies.[1] The drafts appear in your Superhuman drafts folder, and you review and send them yourself. As long as send_draft is set to confirmation in the client, the agent sends nothing without your yes. If it is set to "Always Allow" in Claude, it can send.[1]
In Command Mail the same pass runs on mail:read and mail:draft.[3] According to the documentation, if the connection has only these two permissions, the agent cannot send, even if someone switches off every confirmation in the client; with mail:send-approved only the revision you approve would go out.[3][4] If you later want it to answer routine requests itself, you move one mailbox to auto and grant mail:send; the daily limit and the send key remain as brakes.[3] The other mailboxes keep their level.
Which approach fits depends on where you want the limit to sit. The Superhuman help center describes tool approvals in the client, set per tool, and a read-only mode.[1] Command Mail additionally documents permissions per connection and a level per mailbox.[4]
07When Superhuman is the better choice
Superhuman fits if you want an email client for Gmail or Outlook with built-in AI and your agent should handle email and calendar together. Scheduling, availability search, read statuses and team comments are part of the Superhuman MCP; the Command Mail documentation does not mention them.[1][4]
It also fits if you already pay for Business or a sales team works with HubSpot or Salesforce.[2] MCP access then costs nothing extra, and your agent works directly in the accounts you already have, without forwarding. Ready-made skills for the morning briefing, the end-of-day wrap-up and batch drafting come on top, so you do not have to write those routines yourself.[1] If you keep send_draft on confirmation permanently, you have a clear line.
08When Command Mail fits
Command Mail fits if your agent should work with fixed permissions that do not depend only on a client's settings. You can let a sorting agent read but not send, give a second agent drafts only, and move a single mailbox to auto without opening up the others.[3]
It also fits if your addresses are not only at Gmail or Outlook, for example at iCloud, Yahoo or on your own domain. By forwarding, they all arrive in one place.[4] And if you want to try it first, you can start free with reading and drafts.[3]
09Who Command Mail is not for
If you need a calendar, an email client with built-in AI or CRM integrations, Superhuman is the right choice.[1][2] Command Mail does not replace your existing account: forwarding is not sync, and changes do not flow back to the original mailbox.[4]
Sending requires Pro.[3] If you have strict data-location requirements, check the operator and hosting details first.[5]
10Switching and setup
To Superhuman: you need a Business plan, Ask AI and the Chrome extension for sign-in. In Claude Code you add the server with claude mcp add --transport http superhuman-mail https://mcp.mail.superhuman.com/mcp and authenticate under /mcp.[1] Then set send_draft to confirmation in your client.
To Command Mail: connect each address in the app and set up forwarding at your current provider to the private inbound address Command Mail shows for it.[4] Then connect the agent, in Claude Code for example with claude mcp add --transport http commandmail https://commandmail.app/mcp.[3] Give it only mail:read and mail:draft at first and keep the mailboxes on suggest.
Running both gives you two places to work. Decide up front where your agent sorts and replies, so that archive and drafts do not drift apart.
11Further reading
- Other AI email clients compared
- Permissions and autonomy in Command Mail
- Human in the loop for email agents
You can start Command Mail for free and let your agent read only at first.
Frequently asked questions
Does Superhuman have an MCP server?
Yes. The Superhuman Mail MCP Server is hosted at https://mcp.mail.superhuman.com/mcp and connects Claude, ChatGPT, Codex, Cursor, Gemini CLI, Perplexity and other clients to your inbox and calendar through OAuth. It requires the Business plan or higher and is not available on Starter.
Can an agent send email through the Superhuman MCP?
Yes. The send_draft tool can send immediately or schedule the send, including Smart Send, and undo_send recalls a message within the undo window. Superhuman advises turning on confirmation for every send_draft call in your AI client, because otherwise the email goes out instantly.
How does Superhuman pricing compare to Command Mail?
Superhuman Starter costs $30 per month or $300 per year, Business with the MCP server $40 per month or $396 per year. Command Mail starts free but cannot send on the Free plan; Pro costs €19 per month for the first 100 paying Pro accounts while the subscription runs, and €29 for later accounts.
Does Superhuman work with every email provider?
The MCP server connects Gmail and Outlook accounts that sit on your own Superhuman seat. Command Mail accepts mail from any provider by forwarding, but does not sync changes back to the original mailbox.
Can I use Superhuman and Command Mail together?
Yes, if you also forward a Gmail or Outlook address to Command Mail. You then have two separate states: what the agent archives in Command Mail stays unchanged in Gmail and therefore in Superhuman.