Skip to content
SecureThread
Contact salesSign inGet started
SecureThread AI

Writes with you. Never sends without you.

Draft from one sentence, rewrite with one tap, check before it goes, reply from the inbox and read long mail in four points. Every suggestion waits for you, on the model your organization chose, with contact details and secrets kept back.

Get startedSee it write
  • Suggestions only
  • Your model and key
  • Secrets never sent

What it does

What you use every day

Each one is a button in the app, with the same name as here. None of them sends anything. Every AI feature, with its limits and what it shares, is listed in full further down.

Draft from a sentence

Describe the email and SecureThread AI writes it, with a subject and preview line, in your organization’s design. Anything it was not told becomes a [placeholder] and comes back as a question.

Composer › Write with AI

Rewrite with one tap

Polish, Shorter, More formal, Friendlier, More confident or Apologise. For designed emails it edits the text blocks only, so the layout, signature and footer stay exactly as they were.

Composer › quick changes

Subject lines

Five subject lines for the email in front of you, to pick from or ignore.

Composer › Subject lines

Strengthen the ask

Rewrites the call to action so the reader knows exactly what to do next and by when.

Composer › Strengthen the ask

Check it before it goes

Reads the draft for missing information, an unclear call to action, inconsistent details, tone, unsupported claims and placeholders left in, then offers to fix them.

Composer › Check it

Fill a template’s details

Tell it the facts once, in your own words, and it fills the template’s variables. Anything you did not give it stays empty for you to complete.

Composer › Details

Help me write a reply

In the inbox, say what you want to answer and it writes the reply paragraphs into the editor, reading the message it answers from your mailbox as it is now.

Inbox › Help me write

Summaries of long mail

One to four short points, most important first, above long messages. Images and PDFs attached to the message are read too when your model supports them.

Inbox › Summary

Security scores

Each incoming message gets a score from 0 to 100, higher being safer: Looks safe, Be careful or Likely dangerous, with the reasons in plain words. Failed sender checks hold the score down.

Inbox › Security score

Every AI feature, in detail

In the composer

One email, five ways

Choose a quick change and the same email comes back rewritten. In a designed email it rewrites the text blocks only: your layout, signature and footer never move. You see the suggestion beside your current version and keep whichever you prefer.

Try it: choose a change. The example is written for this page.

SubjectInterview next weekCurrent

Hi Casey, we would like to talk to you about the designer role. Can you do Tuesday or Wednesday next week? Let me know what works. Thanks, Maya

Your words, as you wrote themApply to draft

Before you send

Check it catches what you missed

Missing information, an unclear call to action, details that disagree with each other, the wrong tone, claims nothing supports, and placeholders still in the text. Choose Fix these for me, or fix them yourself.

  • Link left outA link that was not in your email or instruction is taken out.
  • Check this detailFigures or dates you did not supply are pointed out.
  • New contact detailAn address or number it introduced is flagged.
  • Contact detail left outA hidden contact detail that landed inside a link or attribute is dropped and named.
  • Personalisation changedA template variable it touched is shown.
  • Unsafe markup removedUnsafe HTML in the answer is removed.
  • May be incompleteAn answer that stopped early, or parts that could not be used, are named.

Every suggestion is checked before you see it, and carries these warnings when they apply.

In the inbox

Read faster. Spot the scam.

Long messages get a summary of up to four points, and every incoming message a security score from 0 to 100, with the reasons in plain words.

Help me write

Say what you want to answer; the reply paragraphs land in the editor, written from the message as it is in your mailbox now.

Attachments, read

Images and PDFs attached to a long message are read for its summary when your model supports them: up to five files, 20 MB in all.

Held from AI

Mail that tells an assistant to drop its instructions or poses as a system message is kept out of summaries, replies and scores until someone releases it.

Privacy

What leaves, and what never does

Before any text goes to your model provider, passwords, codes and keys become placeholders, always. Email addresses and phone numbers do too while Redact contact details is on, which it is unless an administrator turns it off. The real values come back only in the suggestion you see.

  • Nothing is kept by default: each ask leaves a record of sizes and outcome, never the words.
  • Conversations and memory are off until an administrator turns them on, and only their owner can read them.
  • Conversations, memory and usage records are cleared after your retention period, 90 days unless you choose another.
  • SecureThread trains no models on your mail.

In your draft

Hi Alex, your account is ready. Sign in with the temporary password Harbour-7731 and change it on first use. Questions? Write to [email protected] or call +1 555 0142.

Sent to the AI provider

Hi Alex, your account is ready. Sign in with the temporary password [[SECRET_01]] and change it on first use. Questions? Write to [[CONTACT_01]] or call [[CONTACT_02]].

The suggestion you see has the real values back. The provider never had them.

Your model, your rules

Bring your own model and key

Administrators choose the provider, the model and five switches in Administration › AI drafting. The key is encrypted in SecureThread’s vault and never shown again.

Anthropic

Claude models through Anthropic’s own SDK, streamed. Enter the model ID, such as claude-opus-5-5 or claude-sonnet-5-5, and your key.

OpenAI-compatible

Any chat-completions endpoint over HTTPS. Every address it resolves to must be public, and SecureThread connects only to the address it checked.

Test sends one tiny probe prompt, never mail, and records when the provider last answered. Choose effort, the output limit and whether the model reads images and PDFs.

AI drafting switches

  • AI drafting

    SecureThread AI is available to people whose role allows it. Off, every AI control disappears.

    Starts off

  • Content may be processed by the AI vendor

    An administrator acknowledges that drafts and messages go to the model provider you chose before anything is sent there.

    Starts off

  • Redact contact details

    Email addresses and phone numbers become placeholders before content leaves, and are put back only where the answer keeps them.

    Starts on

  • Keep conversations

    Each person’s asks and answers are kept as their own history, which only they can see and delete.

    Starts off

  • Remember what people teach it

    Lasting facts from a person’s instructions, up to 50 each, carried into their later asks. They can read and delete every one.

    Starts off

Use cases

What teams do with it

Examples of real buttons used by different teams. The instructions in quotes are examples.

Hiring

Offer letters from your notes

Pick the offer template, type the role, start date and salary as you would say them, and Details fills every variable. Check it catches the start date you forgot.

“Product Designer, starts 3 March, salary as agreed, reports to Priya”

Sales

Follow-ups that ask for something

Draft the follow-up from one line, then Strengthen the ask so the meeting request is specific, and choose from five subject lines.

“Follow up with Alex after Tuesday’s demo, suggest a pilot”

Support

Replies without starting from blank

Help me write drafts the answer to the message you are reading; the summary above a long complaint tells you what it is about before you open it.

“Apologise for the delay and confirm the refund is on its way”

Finance

Fewer bank-detail scams

Security scores flag mail that asks for a payment or a change of bank details, or that pushes you to act urgently, before anyone acts on it.

Likely dangerous · “It asks for a payment or a change of bank details.”

Leadership

Announcements in the right voice

Write the update once, then switch between Concise, Confident and Formal until it sounds like you, without touching the design.

“Announce the new office opening to all staff, warm but brief”

AI teams

Assistants that prepare, people that send

Claude, ChatGPT or Cursor prepare drafts through the MCP server; every send they ask for waits for the person behind them, unless your policy allows a narrow automatic send.

Every send needs approval · the default for new agents

Agents and MCP

Your own assistants, on a short lead

Claude, ChatGPT, Cursor and other assistants connect through SecureThread’s MCP server as one signed-in person, with that person’s permissions. Give each one an agent identity with an owner, a policy and one of four modes.

  • Drafts only

    The agent prepares drafts; people send them.

  • Every send needs approval

    The default for new agents: each send waits for its owner.

  • Sends within policy, tells its owner

    Narrow sends go out, and the owner hears about each one.

  • Sends within policy

    Narrow sends go out inside the agent’s limits.

AI agentsMCP server

Reference

Every AI feature, in detail

What each one does, where it is in the app, who can use it, what is sent to your model provider, what SecureThread keeps, its limits and who stays in control.

Composer

Composer

Drafting assistant

A chat beside your email that writes new drafts and changes the one you have, as suggestions you apply or ignore.

  • New draft: describe the email and it writes it with a subject and preview line. Starter ideas: Meeting invitation, Follow-up, Thank you, Announcement, Request, Apology.
  • Tone for a new draft: Natural, Friendly, Formal, Concise, Confident or Apologetic.
  • With no layout chosen it writes into your organization’s standard card design, or into a template’s layout if you picked one.
  • Details it was not given become [placeholders] and come back to you as questions.
  • Edit: type any change in your own words, and keep chatting; it remembers up to 20 earlier turns.
  • Suggestion and Current side by side, previewed in a sandboxed frame; Apply to draft puts it in.
  • On a phone, a new empty email offers “Let SecureThread AI draft this email”.
Where it is
Composer › Write with AI, in a sheet titled SecureThread AI. Controls: New chat, Stop.
Who can use it
People whose role includes AI drafting, once an administrator has switched AI drafting on and chosen a model.
Sent to the model provider
Your instruction, the current subject, the email’s text (for designed emails, only its numbered text blocks) and the earlier turns of the chat, never the rest of the email around them. Secrets and, while redaction is on, contact details are replaced first.
What SecureThread keeps
A usage record (sizes, tokens, outcome). The conversation only if Keep conversations is on.
Limits
30 asks per person every 5 minutes; an instruction of up to 4,000 characters; an email of up to 100 KB.
Who stays in control
Nothing reaches the draft until you apply it, and nothing is sent. Emails written with its help are marked as AI-assisted.

Composer

Quick changes

One-tap rewrites of the email you are writing.

  • Polish, Shorter, More formal, Friendlier, More confident and Apologise.
  • In a designed email it rewrites the text blocks only: layout, images, buttons, signature and footer do not move.
  • Each change comes back as a suggestion beside your current version.
Where it is
Composer › SecureThread AI, the row of quick changes under the chat.
Who can use it
As for the drafting assistant.
Sent to the model provider
The email’s text blocks and the change asked for.
What SecureThread keeps
A usage record only, unless conversations are kept.
Limits
Counted with the composer’s 30 asks every 5 minutes.
Who stays in control
You choose whether to apply each suggestion.

Composer

Subject lines and Strengthen the ask

Five subject lines to choose from, and a sharper call to action.

  • Subject lines: five options for the email in front of you; pick one or keep yours.
  • Strengthen the ask: rewrites the call to action so the reader knows what to do next and by when.
Where it is
Composer › SecureThread AI › Subject lines; Strengthen the ask.
Who can use it
As for the drafting assistant.
Sent to the model provider
The email’s text and current subject.
What SecureThread keeps
A usage record only, unless conversations are kept.
Limits
Counted with the composer’s 30 asks every 5 minutes.
Who stays in control
A chosen subject or rewrite goes in only when you apply it.

Composer

Check it

A read-through of your draft before it goes, with an offer to fix what it finds.

  • Looks for missing information, an unclear call to action, details that contradict each other, tone, claims nothing supports, and placeholders left in.
  • Fix these for me turns the findings into a suggested rewrite.
Where it is
Composer › SecureThread AI › Check it.
Who can use it
As for the drafting assistant.
Sent to the model provider
The email’s text and subject.
What SecureThread keeps
A usage record only, unless conversations are kept.
Limits
Counted with the composer’s 30 asks every 5 minutes.
Who stays in control
Findings are advice; nothing changes until you apply a fix.

Composer

Fill a template’s details

Tell it the facts once, in your own words; it fills the template’s variables.

  • Works on the template you chose, reading its variables and what each one is for.
  • Values you did not give stay empty for you to complete, so nothing is made up.
Where it is
Composer › SecureThread AI › Details.
Who can use it
As for the drafting assistant.
Sent to the model provider
Your instruction and the names and descriptions of the template’s variables.
What SecureThread keeps
A usage record only, unless conversations are kept.
Limits
Counted with the composer’s 30 asks every 5 minutes.
Who stays in control
You review the filled details before applying them.

Inbox

Inbox

Help me write (replies)

Reply paragraphs for the message you are answering, written into the reply editor.

  • Say what you want to answer; the paragraphs appear in the editor for you to edit.
  • The message being answered is read from your mailbox as it is now and treated as untrusted content.
  • Press / in the reply editor to open the bar; Stop ends an answer early.
Where it is
Inbox › reply editor › Help me write, then Ask SecureThread AI.
Who can use it
People who can read the mailbox and whose role includes AI drafting. Not available through an administrator’s access to someone else’s mailbox.
Sent to the model provider
The message text and subject, its image and PDF attachments when the model reads them, and your instruction, with secrets and (while redaction is on) contact details replaced.
What SecureThread keeps
Nothing beyond a usage record, unless Keep conversations is on.
Limits
60 per person every 5 minutes. Messages held from AI are refused.
Who stays in control
Nothing is sent without your normal reply step.

Inbox

Summaries of long mail

One to four short points above a long message, most important first.

  • Appears on messages of 50 words or more: Summarising this message, then Summary ready.
  • Images and PDFs attached to the message are read too when your model supports them; other files are only named.
  • Hide summaries and Show summaries are remembered on your device.
  • Each summary says it was written by SecureThread AI from this message, and to check important details.
Where it is
Inbox › the reader, above the message.
Who can use it
People who can read the mailbox and whose role includes AI drafting.
Sent to the model provider
The subject, the sender’s display name, the message text (up to 40,000 characters) and eligible attachments.
What SecureThread keeps
Nothing beyond a usage record, unless Keep conversations is on.
Limits
60 per person every 5 minutes; up to 4 points of 300 characters; up to 5 attachments, 10 MB each, 20 MB in all.
Who stays in control
A reading aid: opening a summary does not mark the message read.

Inbox

Security scores

A score from 0 to 100 for every incoming message, higher being safer, with the reasons in plain words.

  • Levels: Looks safe from 75, Be careful from 40, Likely dangerous below 40; Checking security while it works.
  • Reasons such as: links go to unusual or unrelated websites; it asks you to sign in or share a password or code; it asks for a payment or a change of bank details; it pushes you to act urgently.
  • Sender checks set a ceiling on the score, so a message from a forged or failing sender cannot score as safe, however it is worded.
  • Shown as a chip in the message list and a panel in the reader.
Where it is
Inbox › message list and reader › Security score.
Who can use it
People who can read the mailbox and whose role includes AI drafting.
Sent to the model provider
The subject, sender name and address (only the domain while redaction is on), reply-to addresses, the sender-check results, the start of the text, its links and the names and types of attachments.
What SecureThread keeps
The score and reason codes for 90 days, never the mail.
Limits
40 requests per person every 5 minutes, up to 25 messages each.
Who stays in control
Advice only: the score never moves, deletes or blocks a message.

Inbox

Mail held from AI

Messages that try to manipulate an assistant are kept away from it.

  • Incoming mail with text aimed at an AI assistant, role-play or fake system messages, or hidden content is graded by SecureThread’s inbound screening.
  • Mail graded high is held from summaries, reply drafts and security scores until someone who can read it releases it.
  • Release and undo are per message, and both are recorded in the audit log.
Where it is
Inbox › screening banner on the message.
Who can use it
Applies to everyone; anyone who can read the message can release it.
Sent to the model provider
Nothing: screening is pattern matching inside SecureThread, with no model call.
What SecureThread keeps
For a released message, a hashed key of the mailbox and message, never the message itself.
Limits
Only messages graded high are held.
Who stays in control
People decide what the assistant may read.

Outgoing mail

Outgoing mail

AI review of outgoing mail

An optional second check after the built-in word list, for abuse that has no swear words.

  • Reviews the subject, preview, sender name and body of each screened send for profanity, insults, harassment, threats, hate and sexual content, at mild, moderate or severe.
  • A finding must quote text that is really in the message, or it is discarded.
  • Hold messages when the AI review can’t run, if you prefer waiting to sending unchecked.
  • Saving a change runs a test review first.
Where it is
Administration › Content policy › AI review (Turn on AI review).
Who can use it
Administrators who manage settings switch it on; it then applies to every screened send.
Sent to the model provider
The subject, preview, sender name, body (up to 40,000 characters) and the alt, title and label text of its images. Card numbers, identity numbers and secrets are masked first; contact details are replaced while redaction is on.
What SecureThread keeps
For a flagged message: the finding, a masked excerpt and the model’s name.
Limits
Waits up to 25 seconds per message.
Who stays in control
A flagged send is stopped and counted under the content policy; the hold switch decides what happens when the review cannot run.

Assistants and agents

Assistants and agents

MCP server for your own assistants

Claude, ChatGPT, Cursor, VS Code and other assistants work in SecureThread as one signed-in person.

  • A Model Context Protocol server at your SecureThread address, /mcp.
  • Tools for drafts and sending, sent-message status, templates, contacts, scheduled sends, campaigns, workflows, analytics, approvals and more.
  • Each connection asks “Allow this app to use SecureThread as you?” and shows the access levels it wants.
  • Administrator access needs a session verified with two-step sign-in and lasts at most 14 days.
  • SecureThread runs no model for these calls: the assistant is yours.
Where it is
Administration › AI access: connected assistants, tool list, Check connection, Create token, Revoke, Disconnect.
Who can use it
People whose role includes using AI assistants, once an administrator switches AI access on.
Sent to the model provider
Tool results go to the assistant you connected, within your permissions; nothing goes to SecureThread’s AI provider.
What SecureThread keeps
Tokens as one-way hashes. Each tool call is audited with its name and outcome, never its arguments.
Limits
Per 5 minutes: 600 requests per connection, 900 per person, 300 tool calls per connection. A connection lasts at most 90 days.
Who stays in control
Disconnect or revoke ends access at once; an administrator can revoke any connection.

Assistants and agents

Send requests and automatic sending

When an assistant wants to send, a person approves the exact email first.

  • Each request holds the exact recipients, subject, message, attachments and mailbox, and lists why it needs approval.
  • Only the person whose assistant asked can Approve and send, or Decline. Approval runs every send rule again.
  • Requests expire after 24 hours. Scheduled sends and workflow runs an assistant starts always wait for a person.
  • Allow automatic sending (off by default) lets narrow emails go without approval: straight from a shared template (by default only a published one), variables only, from the person’s own mailbox, no attachments, within the recipient limit (1 by default, 50 at most) and, by default, only to people in the organization.
Where it is
Administration › AI access › Sending without approval; requests listed under AI access.
Who can use it
Administrators set the policy; each person approves their own assistant’s requests.
Sent to the model provider
Nothing to an AI provider.
What SecureThread keeps
The request until it is decided or expires; every request, approval, decline and automatic send is audited.
Limits
Requests expire after 24 hours.
Who stays in control
A person approves every send unless the narrow automatic rule applies.

Assistants and agents

AI agents

A named identity for each assistant, with an owner, a policy and a mode.

  • Modes: Drafts only; Every send needs approval (the default); Sends within policy, tells its owner; Sends within policy.
  • A policy can only narrow the organization’s rules: allowed recipients (up to 200), a daily send limit, mailboxes, templates, and whether only published templates may be used.
  • Pause stops an agent at once; Revoke is permanent. Each needs a reason.
  • Activity shows what the agent did, never message bodies.
Where it is
Workspace › Agents: New agent, Pause agent, Revoke agent.
Who can use it
On plans that include agents; the number of agents follows the plan.
Sent to the model provider
Nothing to an AI provider.
What SecureThread keeps
The agent, its policy and its activity; every change is audited.
Limits
The daily send limit is a rolling 24 hours.
Who stays in control
The owner is accountable, and administrators set the most any agent may do.

Controls

Controls

Redaction, history and memory

What leaves for the provider, and what SecureThread keeps about your asks.

  • Secrets are always replaced before text leaves: values after labels such as password, PIN, verification code, API key or token, and well-known key formats.
  • Redact contact details (on by default) replaces email addresses and phone numbers with placeholders, put back only in what you see.
  • Keep conversations (off by default) keeps each person’s asks and answers as their own history: composer chats, summaries and reply drafts.
  • Remember what people teach it (off by default) keeps up to three lasting facts per ask, 50 per person, which each person can read and delete on their AI history page.
  • Conversations, memory and usage records are cleared after the organization’s retention period, 90 days unless an administrator sets another; security scores are kept for 90 days whatever the setting.
Where it is
Administration › AI drafting › switches; each person’s AI history page.
Who can use it
Administrators set the switches; each person manages their own history and memory.
Sent to the model provider
Only the placeholder versions of secrets and, while redaction is on, of contact details.
What SecureThread keeps
By default, usage records only: sizes, tokens and outcomes, never words.
Limits
Memory: 300 characters a fact, 50 facts a person.
Who stays in control
Only the person can read their history and memory, and delete one item or all.

Controls

Models, keys and usage

Your organization chooses the model and brings its own key.

  • Anthropic’s Claude models, or any OpenAI-compatible endpoint over HTTPS.
  • Set the model ID, effort, the output limit (256 to 64,000 tokens) and whether it reads images and PDFs. Use this provider switches it on or off; Make this the default provider picks the one asks go to first.
  • Test sends one tiny probe prompt, never mail.
  • Every answer is checked: invented links are removed, and new facts, new contact details, dropped placeholders, changed variables and unsafe HTML are flagged.
  • A usage card shows asks by operation and outcome for any period up to a year.
Where it is
Administration › AI drafting.
Who can use it
Administrators who manage AI.
Sent to the model provider
The key goes only to the provider it belongs to.
What SecureThread keeps
The key, encrypted in SecureThread’s vault and never shown again; usage records for the retention period.
Limits
An OpenAI-compatible endpoint must resolve to public addresses only.
Who stays in control
Switch AI drafting off and every AI control disappears at once.

Limits

Generous, and written down

The limits the app enforces, per person, so one busy afternoon cannot run up a bill or slow everyone else.

Composer asks

Limit

30 per person every 5 minutes

Notes

Drafts, rewrites, subject lines, checks

Inbox summaries and reply drafts

Limit

60 per person every 5 minutes

Notes

Each

Security scores

Limit

40 requests per person every 5 minutes

Notes

Up to 25 messages a request

Email size for AI

Limit

100 KB

Notes

Longer emails are refused with a plain message

Your instruction

Limit

4,000 characters

Notes

Plus up to 20 earlier turns of the chat

Summary

Limit

Up to 4 points

Notes

300 characters each; text past 40,000 characters is not sent

Attachments read

Limit

Images and PDFs, up to 5 files

Notes

10 MB each, 20 MB in all; inbox summaries and reply drafts

Questions

About SecureThread AI

The full guide, with every switch and limit, is in the documentation.

app.securethread.co/compose
The composer: a message being written from a template, with a live desktop and phone previewThe composer: a message being written from a template, with a live desktop and phone preview
Does SecureThread AI ever send email by itself?

No. In the composer and the inbox it only suggests: nothing reaches your draft until you choose Apply to draft, and nothing is sent until you send it. Emails written in the composer with its help are marked as AI-assisted. External assistants connected through MCP ask before every send unless an administrator allows automatic sending for narrow, template-only emails.

Which AI models does it use?

The one your administrator chooses. SecureThread works with Anthropic’s Claude models, and with any OpenAI-compatible endpoint over HTTPS. Your organization brings its own API key, which is encrypted in SecureThread’s vault and never shown again.

What is sent to the AI provider?

Only what the task needs: the email or message text, your instruction, the subject and, in a chat, up to 20 earlier turns. Passwords, codes and keys are always replaced with placeholders first, and email addresses and phone numbers are too while Redact contact details is on. Attachments are read only for inbox summaries and reply drafts, and only images and PDFs the chosen model can read.

Does SecureThread keep what I asked?

Not by default. Each ask leaves a usage record with sizes, token counts and the outcome, never the content. Conversations and the assistant’s memory are kept only if an administrator switches them on, and then only the person who asked can read or delete them. Conversations, memory and usage records are cleared after your organization’s retention period, 90 days unless you set another. Security scores are kept for 90 days whatever the setting.

Can it be tricked by an email I receive?

Mail that tells an assistant to drop its instructions, poses as a system message or asks it to send or reveal data is held from summaries, reply drafts and security scores until someone who can read it releases it. Milder signs, such as hidden text, are flagged to the model as a caution. Messages are always treated as untrusted content, and every answer is checked: links it invented are removed and new facts are flagged.

Who can use it?

Anyone whose role includes AI drafting, once an administrator switches it on for the organization. Administrators choose the model, the switches and the retention period, and see usage by operation.

Turn it on for your workspace

An administrator adds a model and key, switches on AI drafting, and every person whose role allows it sees Write with AI in the composer.

Plans with AI

Send your first approved email today

Connect a mailbox, pick a template and send with the checks and approvals your organization needs. No new email address, no new habits.

Get startedTalk to us
Send from the mailboxes you already useA person approves what needs approvalCredentials encrypted with AES-256-GCMEvery action in an append-only audit logAI assistants governed over MCP1,499 starter templatesHonest status: accepted is never called delivered