Your data deserves
a conversation.
Ask the way you talk. dulyx queries the ERP, CRM or API your company already has, answers with live data and only writes anything after your yes.
- Duly assistant
- Léo finance
- Sofia support
- Íris logistics
scheduled workflow
The summary is already in.
Nobody has opened the ERP yet. A workflow ran at the set time, pulled the stalled orders and sent the summary to the manager's WhatsApp, before the meeting.
illustrative scenario · numbers from the demo ERP
- 1Triggerevery day · 08:00
- 2Search Ordersstatus = under review
- 3Conditionany stalled?
- 4Notifymanager's WhatsApp
Good morning. 4 orders under review, R$ 123.00 in total. The oldest is from 07/20, stalled for 43 days. Want me to release any of them?
scheduled · every day at 08:00
I'll check in the meeting, thanks
the usual way
Without AI, four steps.
To find the latest order under review: open the ERP, filter, sort and open the order. Every day, for every question, by whoever knows where to click.
real turn · demo ERP
and again tomorrow, for every question
Renata · sales representative
Which customer stopped buying?
39 customers in Belo Horizonte. To find who stopped, she would export 2,649 orders and compare dates, customer by customer. Over an hour, and whoever stopped would slip by unnoticed.
illustrative scenario · customers from the demo ERP
- 1export the 2026 orders2,649 orders
- 2filter the city: Belo Horizonte39 customers
- 3sort by customerA → Z
- 4compare dates, one by one0/39
whoever stopped shows up in no report
which BH customers haven't bought in 45 days or more?
17 of the 39 BH customers have gone 45 days or more without buying.
- Olívia Gomes #271last purchase 04/15 · 3 orders this year138 days
- Wagner Klein #26810 orders this year · R$ 18,989.3272 days
- Lucas Junqueira #68 orders this year · R$ 17,875.8363 days
on your company site
What does Dulyx actually do?
It puts an agent inside your system, through the API it already has.
Hi! I can check price, stock and delivery times, and handle orders right here.
do you have 2kg brown rice?
We do, $6.20. Want me to add it to your cart?
add 2 please
Done. Your cart has 2 × Brown Rice 2kg, total $12.40.
on the company WhatsApp
Does it work on WhatsApp?
The same agent, on your company’s number.
send me the open orders every day at 8am
Got it: every day at 08:00, on this number. Since it’s recurring, can you confirm?
yes
scheduled · every day at 08:00tomorrow3 open orders. #45892 Carter, 2 days ago · #45901 Bennett, today · #45907 Foster, today. Want me to release any?
- asked in chat
scheduled every day at 08:00, next send tomorrow
- outside 24 h
Meta-approved template · R$ 0.035 per message
- a teammate replied
the agent goes quiet in that chat for 30 min
WhatsApp Business Platform, through Meta’s official API.
Embedded Signup on your company’s number. No virtual number, no third-party app and none of the bans that take down unofficial automation.
- Replying to whoever wrote in
- R$ 0within the 24 h conversation window
- Message outside the conversation
- R$ 0.035per delivered message · billed by Meta to your account, Brazil rate
- Dulyx markup
- R$ 0just the agent’s credits, when it replies
in Claude, ChatGPT or Cursor
Who calls my API?
Dulyx does.
The model doesn’t.
what’s the latest order under review?
It’s order #5996, from Úrsula Oliveira (Porto Alegre/RS), placed on 08/30/2026, totaling R$ 1,250.96.
release it
Before writing: order #5996 goes from A (under review) to L (released). Confirm?
yes
Done, #5996 is released.
- Reaches the model
The question, the agent’s instructions and the slice of data that survived the cut.
- Never reaches it
URLs, credentials, database access and anything not enabled in the context.
- the same connector
Works the same in ChatGPT, Cursor and any client that speaks MCP.
on the team board
Does it organize the team’s work?
Boards the agent reads, creates and moves from the chat.
To do3
Doing2
Done2cards here count as done
create a card to call the supplier by friday
Created in To do, due Friday, owner you. I attached the quote that was in the conversation.
where the data goes
The model asks. dulyx executes.
On the AI side, only question and answer in text. Credentials, URL and request stay on the dulyx side, and a cheap model cuts what the expensive model doesn't need to see. What isn't sent isn't billed.
real turn · demo ERP
- Context resolvedGESTAO_PEDIDOS
16 ms - GET /api/pedidos?status=Athe caller was dulyx
200 - Tax IDs (CPF, CNPJ)masked before going up
ok - Data cut5,939 read → 564 sent
−90.5% - PUT /api/pedidos/5996confirmed by you
200 - Verificationa second model checked the answer
ok
before the end-of-day meeting
Can it put together reports?
Pages, PDFs and spreadsheets, with live data.
Open orders · September 2026
| order | customer | amount | days |
|---|---|---|---|
| #45892 | Carter | $12,480 | 2 |
| #45901 | Bennett | $3,920 | 0 |
| #45907 | Foster | $8,150 | 0 |
| #45880 | Sloan | $1,275 | 5 |
guided mode
What if it wants to change something?
Nothing is written without your yes.
In chat · ordersokGET /pedidos?status=AokIf due date < todayokAwaits approvalwaiting for youmove 3 orders to collections
ApproveRejectPUT /pedidos/statusTable in chatand again tomorrow
And what does it do without me asking?
It runs on the schedule you set.
- every day · 07:30Stalled orders4 steps · WhatsApp alertTrigger→Action→Condition→Alertok · today 07:30
- Mon to Fri · 18:00Daily cash close6 steps · runs on its ownTrigger→Action→Loop→Actionok · yesterday 18:00
- every 15 minStock below minimum3 steps · WhatsApp alertTrigger→Action→Condition→Alertskipped · API down
- day 1 · 09:00Overdue invoices this month3 steps · table in chatTrigger→Action→Displaynext · day 1 · 09:00
- in chatRelease order5 steps · asks for approvalTrigger→Action→Approval→Action12 runs · last one today
who answered today
Not an assistant. A team.
Each agent is a role: a name, a way of answering and a clear limit on what it can see, do and spend. A platform rule, not a prompt.
Duly Assistant
Answers about the data and documents you registered. Ready out of the box in a new account.
readdocumentsLéo Finance
Invoices, overdue bills and duplicate copies. Read-only: changes require a person.
read onlySofia Support
Manuals, procedures and FAQ, in the site chat and on WhatsApp, with the same memory.
readsite + WhatsAppÍris Logistics
Tracking and delivery changes. Every change goes through confirmation.
write with confirmation
In the dashboard, on the company WhatsApp, in your site chat and in Claude, ChatGPT or Cursor, through MCP. The same agent, the same memory, the same limits.
the turn, step by step
How do I know what it did?
Every answer leaves a trail.
security
What are the guardrails?
Mechanical limits, at every layer.
- Credentials stay home
URLs, keys and requests stay on the Dulyx side. The model gets the question, the agent’s instructions and the slice of data.
- Tax IDs masked
Brazilian CPF and CNPJ numbers are detected by value, in any field, and swapped before going up. The real number only returns when your system is called.
- Data funnel
A cheap model cuts what the expensive model doesn’t need to see. In a real turn: 5,939 read, 564 sent.
- Scope per token and per agent
Each token sees only the contexts enabled for it; each agent, only its own. The stricter of the two always wins.
- Writes only with your yes
The proposal shows what will change and where, then waits. It expires: a late yes doesn’t write.
- Caps on rounds, calls and credits
Per turn, including when consulting another agent. Mechanical: it doesn’t judge merit, it just counts.
- Capped pagination
At most 20 pages and 2,000 records per query. A source that repeats a page never turns into an endless loop.
- Blocked destinations
Internal or private addresses are never called, not even through a redirect.
- Checked by a second model
It checks the answer against the turn’s data before sending. That call is on Dulyx.
- Moderation
Insults, discrimination, threats and defamation never go out in your company’s name. Without a verdict, the answer doesn’t go out.
- Credit caps at four levels
Account, context, token and agent. Hit the cap and it blocks. Overage is never charged unless you buy it first.
- A trail for every turn
What was looked up, cut and blocked is recorded. The data used tab exports to CSV, Excel and PDF.
- Human pause on WhatsApp
If someone on your team replies from the app, the agent goes quiet in that conversation for as long as you set.
- Number linked by code
Before scheduling a reminder, the person confirms their own number with a code. Linking doesn’t open a turn or use credits.
- Website chat only on your domains
The public key only answers on allowed domains, with a cap per visitor. Anywhere else, it doesn’t answer.
- Purge on a deadline
Attachments, generated files and session logs are deleted from disk after the number of days you choose.
pricing
How much does it cost?
You only pay for what reaches the model.
the rest of the dashboard
What else does it do?
The rest of the dashboard, one line each.
- Memorylearning
an answer approved in chat comes back the same · no new AI call
- Knowledge basedocuments
PDF, image, spreadsheet and web page · up to 20 MB · AI review
- Environment mapoverview
connected contexts, actions and sources · turns lighting up live
- Agent modesbehavior
with AI · decider, no model in the turn · guided by a script
- Handoffteam
one agent consults another · each within its own scope
- Verificationquality
a second model checks the answer before sending
- Users and permissionsaccess
per person and per agent · no code scattered around
- Asker profileconversation
language, tone and level of detail · sent with every message
- Usage and creditsaccount
balance, cost per conversation and response time · buy without leaving
- Conversationshistory
CSV, Excel and PDF · integration health · data used
- Website chatchannel
bubble, inline, React or standalone link · cap per visitor
- API and streamingcode
scoped token · SSE · cURL, Node, Python and C#
- Files in the conversationattachments
receives files · returns PDFs
- Personal dataLGPD
CPF and CNPJ masked before reaching the model
- Model tiersmodels
you pick the tier, not the model · switches on its own if one goes down
technology
And what does it run on?
Only official APIs, under direct contracts.
- WhatsApp Business PlatformMeta
official Cloud API · Embedded Signup on your number · approved message templates
- OpenAI APImodels
GPT family · direct contract with the official API
- Claude APImodels
Claude family, by Anthropic · direct contract with the official API
- Gemini APImodels
Gemini family, by Google · direct contract with the official API
on video
Can I see it in a minute?
In 57 seconds, a real turn in the demo ERP.
Read instead of watching
- “Qual o último pedido em análise?” (which is the latest order under review?). Without AI, it takes four steps to the ERP: open the ERP, filter by status, sort by date and open the order.
- Dulyx plugs into the other side of the ERP, through the API it already has, with the AI agents inside.
- A real turn in the demo ERP: “ult ped analise” (latest order under review) goes through Dulyx to the ERP and comes back with order #5996, from Úrsula Oliveira, R$ 1,250.96, and its four items. “liberar” (release) becomes a proposal to change the status from A to L, which waits in Dulyx and only reaches the ERP after Confirm — and stays in the log.
- The model asks, Dulyx executes: the credentials stay on its side, and of the 5,939 records read, 564 go up to the model.
- It is not an assistant, it is a team: Duly, Léo, Sofia and Íris come out of Dulyx, each with a role and a scope.
- Your data deserves a conversation. Waitlist open at dulyx.com.br.
getting started
Four steps. Once.
No AI team and no model to pick. The four steps at the start of the day were a routine; these you do once, and the agent is live.
- 01Create the contextthe subject and who can ask
- 02Connect your APIsthe one the system already has, or OpenAPI
- 03Set the permissionswhat each agent sees, does and spends
- 04Turn on the channelsdashboard, WhatsApp, site and MCP
Straight answers
What you might still want to ask:
- Still have a question? .
Where does my data live?
In your system, where it always was. Dulyx queries your API on every question and only stores what you told it to: the documents you uploaded and the conversation log. Your data never trains a model.
Where does my data live?
In your system, where it always was. Dulyx queries your API on every question and only stores what you told it to: the documents you uploaded and the conversation log. Your data never trains a model.
What about LGPD and privacy?
Only what you enabled leaves your system, only the slice that’s needed reaches the model, and every access is logged: what went out, through which agent, for what and when. The report your legal team would ask for is the Data used tab, ready to export, not a separate project.
My system has no API. Can I still use it?
You can start today: files, spreadsheets and web pages become a searchable base, without a single line of integration. Without an API, the agent answers about what you uploaded. To query live data and take actions you need an API, which usually comes from your ERP vendor.
Do I need a developer?
To chat, no: the website chat, WhatsApp and AI assistants don’t need any code. To embed it in your software, yes, but not a new team: whoever already looks after your system can do it with the docs in hand, without learning anything about AI. And upkeep afterwards happens on screen, not in code.
What happens when the AI gets it wrong?
Every answer comes with its source, every change stops for confirmation and, when it isn’t sure, the agent asks instead of making things up. Still got it wrong? The whole conversation is recorded: you find the step, fix the setup and the fix applies from the next answer on.
How much will I pay?
In credits per conversation, with a per-agent limit you set. No per-user pricing and no surprises: the dashboard shows the cost of every answer. Plans and pricing are shared in early access.
What if I want to leave?
Your data never stopped being yours: your system stays the same and revoking a key takes one click. Leaving Dulyx is pulling a plug, not migrating a database.
- Still have a question? .
The conversation starts here.
Early accessAI in your operation shouldn’t cost a whole team.
It should fit in an afternoon, on one screen, on a bill you understand.
Be among the first to put agents inside your system.
What’s your name?
What’s your name?
No spam. Just one note when Dulyx is available.