Solutions

A natural language interface for software you already run

Interfaces have grown faster than people's ability to learn them. A natural language layer changes the unit of interaction from a click to an intention, without replacing the software underneath.

The complexity gap

Enterprise software accumulates capability far faster than it can expose it. Every release adds features, and each one is another item in a menu, another tab, another conditional field. The gap between what the software can do and what an ordinary user can actually accomplish widens with every improvement.

Training closes the gap temporarily and expensively. Redesigning the interface moves the problem rather than solving it, because the capability that has to be exposed keeps growing. A natural language layer is the only approach whose cost does not scale with the number of features, because a sentence describing an outcome is the same length whether the application has forty screens or four hundred.

Mapping intent to execution

The hard part of a natural language interface is not understanding the sentence. Modern language models do that well. The hard part is knowing what the application can actually do, in what order, with what preconditions, and under what permissions - and then doing it reliably.

Merlean solves that with an analysis pass over your source. It builds a model of the endpoints, the fields each one requires, the validation that applies, the sequences that constitute a real workflow, and the decisions a user has to make along the way. Intent is resolved against that model, not against a screenshot or a guess.

Endpoints and their request and response shapes, taken from the code that defines them.
Form fields, types, and the conditional logic that shows or hides them.
Permission checks, so an action is only offered to a user who can perform it.
Multi-step flows, including the order steps must run in and what each one needs from the last.
Decision points, such as choices a user has to make before the workflow can continue.

Three ways users reach it

Embedded chat

A widget inside your product, authenticated as the signed-in user. The most common surface, and the fastest to deploy.

An AI agent

Merlean exposes your application over the Model Context Protocol, so an assistant your users already use can operate it on their behalf.

Your own interface

Call the execution layer directly and build whatever front end you want on top of it - a command bar, a voice interface, a batch runner.

What good looks like

A natural language interface is working when users stop describing steps and start describing results. "Generate an invoice for this customer and send it by email" rather than "open billing, then new invoice, then pick the customer". The first is a goal, and the software should be the thing that knows the path.

It is failing when users have to phrase requests in a particular way to be understood, or when they check afterwards whether it really did what they asked. Both are signs the layer is guessing rather than reading.

Frequently asked questions

Is this the same as a chatbot?

No. A chatbot produces text. A natural language interface produces state changes in your application. The conversation is the input method; execution is the output.

Do we have to rebuild our application?

No. Merlean sits in front of the application you already have and drives it through the interfaces it already exposes. Nothing about how the software works has to change.

What happens when a request is ambiguous?

Merlean asks. Because it knows which fields an action requires and which decisions a workflow contains, it can ask a specific question rather than making an assumption and hoping.

Does it work in languages other than English?

Yes. The intent layer is multilingual, and flow matching uses multilingual embeddings, so a user can describe an outcome in their own language even when the application is in English.

See it against your own application

Merlean reads your repository and shows you what it understood before anything is exposed to a user. The fastest way to judge the fit is to point it at a real codebase.

Get in touch Read the docs