Dulyx is a Brazilian platform that sits between a language model and a company’s internal systems: ERP, CRM, an off-the-shelf SaaS or an in-house API. The model phrases the request; Dulyx checks the scope, calls the customer’s API, filters the result and returns text. The model never receives a URL, a credential, the table layout or permission to call anything.
The problem it solves
A company that wants an agent answering questions about its own operation usually has to build an AI team: choose and version a model, write the tool orchestration, handle per-user permissions, set up observability and keep token costs under control. That is months of work before the first useful answer, and the risk of exposing the internal system to the model is there the whole time.
Dulyx delivers that layer ready to use. The customer registers the context, connects the APIs they already have, generates a scoped key and makes an HTTP call. Before any data goes up to the expensive model, a cheap model discards what is not needed to answer: about 91% of the data read never reaches the model, and what doesn’t go up doesn’t show up on the bill either.
Who it is for
For companies that already have systems running and operational data (orders, inventory, billing, customer service, logistics) and want a person (an end customer or an internal team) to ask and get things done in natural language, without it turning into a platform project. It works both for those with a documented API and for those who only have documents and spreadsheets, and also for those who want their own data inside the assistant their team already uses, via MCP.
How the platform behaves
Every write stops for confirmation: the agent proposes the change, shows the record by name and waits for a person to approve. Every turn leaves a trail: what was queried, what was cut, what was blocked, with time and cost calculated. And the bill comes in credits per conversation, not in tokens nobody knows how to convert.
Stage
The platform is in the final stage of development, with access granted through a waitlist. Already in production: organizations, users and API keys; REST contexts and knowledge base; context routing and chat; credits and execution history; observability; WhatsApp with scheduled reminders; workflows with scheduling; task boards; PDF and spreadsheet reports; agent memory and environment map. Planned: native ERP integrations and official SDKs.
Who builds it
Carlos Corgozinho, Founder & CEO. LinkedIn · ccorgz.com
Contact
The channel is contato@dulyx.com.br, and early-access sign-up is on the waitlist. For technical details, see the developers page.