A marketer, a researcher, developer support, personal assistant, you name it. Every agent has a job, your context, and tools that reach the outside world
Being built now · Beta invites go out first · Nothing sends without your OK
I ordered a drill two weeks ago and it stopped working. Can I return it?
Searched returns-policy.pdf
Power tools carry a 24-month warranty, so this is a warranty claim rather than a return. Reply to your order email with the order number and we'll arrange a replacement.
Built on models from
How it works
Five minutes to set up. A step-by-step guide for each part lives in the app.
Policies, product guides, meeting notes. Your agents answer from them and name the file they used.
Pick one of eight roles and when it works: every Monday at nine, every Friday, or the moment a customer writes.
The work is done and waiting in one inbox. Approve it, edit it, or bin it. A weekly digest sums up the rest.
The work, done
Real drafts from real roles — a researcher, a marketer, an assistant and a support desk, each working from your documents.
Approvals inbox
Every post, email and ticket waits here with its full text. Approve, edit or reject in one click. Every decision is logged.
Nova wants to publish a post
LinkedIn · scheduled run · Monday 09:00
Every tool we sell now carries a lifetime warranty. Not 24 months. Lifetime. Because a drill that quits in year three was never really yours.
Nova · Marketer
Nova reads what changed this week and drafts the post in your voice. You approve it before anyone sees it.
Mia · Support
Mia answers from the returns policy you uploaded and names the file. When it isn't in there, she hands it to you.
Sol · Researcher
Sol runs your standing question every Friday and writes it up with numbered sources you can open and check.
Ada · Dev support
Ada answers the developer from your API docs, reproduces the error, and logs it with the steps to repeat it.
Kai · Assistant
Kai drafts the email from your meeting notes and proposes the next date. It sends when you confirm.
Trust & control
Handing work to an AI is a bigger ask than trying another app. So an agent drafts, and a person decides.
A new agent cannot publish or send anything. You allow that yourself, per agent and per tool.
Posts and emails wait in your inbox with the full text. Approve, edit or reject — from your phone.
Agents answer from your documents and cited web pages. When the answer isn't there, they say so.
Every run, draft and decision is recorded: who did what, and when.
Nothing you upload is used to train an AI model. Saved credentials are encrypted.
Each agent has a rule for handing off. When it applies, the agent stops and brings in a person.
Roles
Pick one and change anything: its job, its voice, the documents it reads, when it asks you.
The difference
ChatGPT and Claude answer quick prompts. Desker gives you autonomous staff that work on your schedule, know your business, and stop for your OK.
| Capability | Generic AI (ChatGPT / Claude) | Desker Autonomous Agents |
|---|---|---|
| Schedule & initiative | Reactive only — sits idle in a browser tab until you manually prompt it | Autonomous 24/7 — runs on schedules, events, and background triggers |
| Company context | Forgotten between sessions or re-pasted into prompt windows repeatedly | Permanent company context, structured profile, and reference documents |
| Taking real action | Outputs text in a chat bubble; cannot reach external tools or workflows | Searches the web, drafts emails, and prepares ready-to-send action items |
| Safety & approvals | No review gate — you manually copy-paste or risk unverified automations | Draft-only by default — every post and email waits for your 1-click OK |
| Job specialization | One generic chat window with no defined role, tools, or boundaries | 8 tailored roles with assigned responsibilities and escalation rules |
FAQ
We're building it now and opening a beta soon. Until then, sign-in is limited to the people testing it.
It answers your clients on a chat link or a widget on your site, searches the documents you upload, researches the web with citations, drafts posts and emails, logs bugs and requests, and schedules its own follow-ups. It can work on a schedule or the moment someone writes in.
No. Every agent starts in draft-only mode, so posts and emails wait for your approval in the Inbox. You can loosen that per agent and per tool once one has earned it.
You catch it in the inbox, because the draft is the thing you approve: reject it, or edit it and send the version you want. Every agent also has an escalation rule that stops it and brings in a person, and the audit log shows what it did and which sources it used.
From the documents you upload (PDF, Word, Markdown, text) and from cited web pages. When the answer is in neither, the agent says it doesn't know rather than inventing one.
Nothing you upload is used to train an AI model, and your documents are readable only inside your organisation. Integration credentials such as API keys are encrypted at rest with AES-256-GCM. It is stored with our hosting and database providers, Vercel and Neon, and deleted when you close your organisation.
Today an agent can send email through your own Resend account, post notifications to Slack or any webhook, take work in through an inbound webhook, and talk to clients through a hosted link or the embed widget. Direct Gmail, LinkedIn and Slack posting are not built yet.
About five minutes: pick a role, upload a document, test it live in the playground, publish. The app has a step-by-step guide for each part, from hiring an agent to reading the audit log.
An agent reads your documents and replies in the language you write its job in, and the models behind it handle the major languages well. The Desker interface itself is English only today.
Desker is still being built. Beta testing opens soon.
No payment during the beta · Nothing sends without your OK