Every request, relayed.

An AI agent works your support inbox: it reads, sorts and answers every mail. Your team stays in the loop — to approve, to correct, or to let it run.

From noise to a desk

Requests come at you from every side.

The agent catches each one, reads it, gives it a topic and an answer — and lines everything up. What still needs a person comes to a person.

incomingclassifieddrafted

The difference

Human in the loop. Not agent in the loop.

Most help desks hand every ticket to a person and let AI help at the edges. relaiDesk turns that around: the agent works every ticket, and people decide where their judgement is needed.

A help desk with AI

MailPersonAI suggestsPerson sends

People do the work, the AI assists. Every mail waits for someone.

relaiDesk

MailAgent reads, sorts, answersPerson decides — if needed

The agent does the work, people keep the say. The routine answers itself.

From drafts to autopilot, one stage at a time

answered by the agent 55 %reaches a person 45 %

Illustration, not a promise: how much a desk hands over depends on your mail and on what you allow. A dry-run switch keeps every mail inside until you say go.

The desk

One conversation per customer. A draft already waiting.

Every mail becomes a ticket with a topic, a language and a suggested reply — with the reasons the agent gave for it.

Features

Everything a support inbox needs. Nothing it doesn’t.

Any model you like

Anthropic Claude, OpenAI, or any OpenAI-compatible API — including models you host yourself. Or none at all: then it is a calm ticket system.

Any mailbox you have

Microsoft 365 and Exchange Online with push, or any IMAP/SMTP account. Replies leave from your own address, threads stay intact.

Every language, both ways

The agent answers in the customer’s language; your team reads every thread — and every sent reply — in its own.

Your knowledge, and only yours

Answers come from a knowledge base with revisions and sources. The agent does not get to invent how your product works.

A pipeline, not a black box

Classify, check against knowledge, draft, decide — in stages, with structured output. Every decision is stored with its reasons.

Guardrails built in

A dry run by default. Auto-replies, out-of-office notes and bounces are never answered. Auto-send only for topics you allow.

See what customers ask

Topics and clusters on a dashboard, and a daily digest mail — so a wave of the same question shows up before it floods you.

A team app on every phone

Roles for owners, admins and agents, passkeys instead of passwords, and push when something needs you.

Self-hosted, white-label

One Docker setup on your server. Your brand on the desk and in every mail, your data in your database.

How it works

Four stages between a mail and a reply.

01 · ingest

Mail becomes a ticket

Threads, attachments and the customer's history, in one place.

02 · classify

Topic, language, urgency

Structured output, checked against a schema — never parsed from free text.

03 · draft

A reply from your knowledge

With the passages it used, so a reviewer sees why in one glance.

04 · decide

Send, ask you, or escalate

By the stage and the rules you set. The desk tells you which one — and why.

Coming soon

Be there when it opens.

Leave your address and we write once — when relaiDesk is on GitHub and ready to install. No newsletter, no sharing.

The list opens in a few days — the form goes live together with it.

Open source

Built in the open. On GitHub soon.

relaiDesk already answers a real support inbox every day. The code is being prepared for release: docs, an example knowledge base and a one-command setup.

AnthropicOpenAI-compatibleMicrosoft 365IMAP/SMTPTypeScriptMongoDBDocker

On the way to the release

  • done:Runs a real support inbox every day
  • done:Licence: AGPL-3.0
  • to do:Documentation
  • to do:Example knowledge base
  • to do:One-command setup with Docker