KoltrixDraft replies with Claude and Koltrix Drafts that sound like you
How create_draft works, how to brief an assistant for a good first draft, make it read the thread first, match your voice, and check a draft before you send it.
On this page(9 sections)
The best use of an assistant in your inbox is not summarizing. It is the first draft of a reply, written from the actual thread, waiting where you can edit it.
Koltrix's MCP server has one tool for this, create_draft, and understanding what it does and does not do is the difference between drafts you can send in thirty seconds and drafts you rewrite from scratch. This guide covers how drafting works, how to ask for drafts that sound like you, and how to check one before it leaves.
What create_draft does
When you ask your assistant for a reply, it calls create_draft. The tool saves an email as a draft in your Koltrix Drafts folder. It never sends. The reply it returns contains a draft id and a review_url, a link that opens the draft in Koltrix where you can read it, edit it and send it yourself.
The inputs are plain:
| Input | Meaning |
|---|---|
body_text |
The message, as plain text. This is the only required input |
to, cc |
Recipient addresses |
subject |
The subject line |
in_reply_to_thread_id |
The conversation this replies to |
from_mailbox |
Which of your mailboxes to send from |
Two behaviors are worth memorizing.
Replies are easy. If the assistant passes the conversation it is replying to, to and the subject default to replying to the latest message. It does not have to guess the recipient, which removes a whole category of mistake.
The From mailbox has a default. If the assistant does not choose one, Koltrix uses the mailbox the conversation was addressed to, or your first mailbox. If you have several addresses, such as support@ and you@, say which one you want, or have the assistant list your mailboxes first. In apps that support interactive cards, a mailbox picker lets you choose.
The draft is plain text. That is a limit as well as a simplification: you will not get a formatted email with images from an assistant, and drafts made this way carry no attachments and no Bcc. For anything richer, open the draft in Koltrix and finish it there.
Ask for drafts the way you would brief a person
Compare these two prompts.
Reply to Dana.
Draft a reply to Dana's latest message in the Acme thread. Say yes to shipping on Friday. Ask for the PO number. Friendly, brief, four sentences at most. Sign off with just my first name.
The first leaves the assistant to invent both the content and the tone. The second gives it facts, a tone and a length. A good brief has four parts:
- The decision. What you are saying yes or no to.
- The facts to include. Dates, numbers, names you want used.
- The ask. What you need back.
- The register and length. Casual or formal, two lines or a paragraph.
If you leave out a fact, the assistant may fill the gap with something plausible. That is the main risk in drafting: plausible and wrong. If you do not know the date, say "do not mention a date" rather than letting it improvise one.
Make the assistant read first
A draft is only as good as the context it came from. Ask the assistant to read the whole thread before writing:
Read the whole thread with Acme first. Tell me in one sentence what Dana last asked, then draft the reply.
The one-sentence recap is a cheap check. If it gets the question wrong, you have caught the problem before reading a draft built on it. Reading a conversation uses get_thread, which returns every message with sender, recipients, date and plain-text body, with attachments listed by name and size. It does not open attachments, so if the answer depends on an attached document, tell the assistant what is in it.
Teach it your voice
Assistants write in a default voice unless you supply yours. Three ways to do that:
- Show, do not describe. "Match the tone of my last three replies to Dana" works better than "be warm but professional," and it gives the assistant something to search for. Style guide for AI-drafted email goes deeper on this.
- Keep a short voice note. A few lines saved in your assistant's custom instructions or in a project: how you open, how you close, words you never use. Paste it when a chat does not have it.
- Edit and say what you changed. "I shortened it and removed the exclamation point. Do this from now on."
The checklist before you hit Send
Drafts are reviewable by design, but only if you review them. A fast check, in order:
- Recipients. Right person, right Cc? A reply defaults correctly, but a new message does not.
- From address. Is this the mailbox you meant?
- Facts. Every date, number and promise: did you give it to the assistant, or did it supply it?
- Tone. Would you say this to this person?
- The last line. Assistants like to offer extra help ("let me know if you need anything else") that you may not want to promise.
- Commitments. Anything in the draft that creates an obligation? You are the one who signs.
What happens to the draft next
The draft sits in Drafts like any other. You can open it from the assistant's link, or from the folder in Koltrix, edit it, schedule it, or discard it. By default the assistant does not send it.
If an admin has allowed assistant sending and you opted in when you connected, there is a second path: the assistant shows you the recipients, subject and full message, you say yes, and it sends that exact draft. Safe sending from Claude: the confirmation flow explains the checks. Many people never turn this on, and drafts you send yourself are a complete workflow.
In apps that show interactive cards, the draft appears as a card with the whole message and buttons to send, edit in Koltrix or discard. Discard only closes the card, and the draft stays in Drafts. See MCP Apps cards.
A realistic loop
| Step | You | Assistant |
|---|---|---|
| 1 | "Which conversations are waiting on me?" | Lists them from the inbox summary |
| 2 | "Draft replies to the first three. Brief, friendly." | Reads each thread, saves three drafts |
| 3 | Open Drafts, edit and send two | |
| 4 | "The third needs a call. Re-draft it asking for 15 minutes Thursday." | Saves a new draft |
| 5 | Send it |
Notice that the assistant did the reading and the first writing, and the human did the judging and the sending. That split is the point.
Mistakes to avoid
- Asking for a draft before the thread is understood. Recap first.
- Letting it decide commitments. Dates, prices and refunds are yours to state.
- Accepting the first draft. The second pass, "shorter," "less formal," "lead with the answer," is cheap and usually better.
- Trusting quoted text. If a draft quotes the customer, check the quote against the thread.
- Forgetting the source is untrusted. Email bodies are text from other people. If a draft includes an instruction it picked up from an email, such as a link or a request that you did not ask for, treat it with suspicion. Defending AI assistants against hidden instructions shows what that looks like.
Key takeaways
create_draftsaves a plain-text draft to your Drafts folder and returns a link to open it in Koltrix. It never sends.- Brief the assistant like a person: the decision, the facts, the ask, the tone and length.
- Have it recap the thread first, and name the From mailbox if you have more than one.
- Review recipients, From address, facts and commitments before you send.
- See the integration overview and the docs for the full tool list.
Start with Koltrix
Your domain, one inbox, and an API that sends.
A team inbox where AI sorts and drafts (nothing is sent without your click), plus the transactional API and SMTP relay your product sends with. 7 days free, no card.
