Skip to content

NexaModus AI

AI on a short leash, with an audit trail.

A single gateway instead of calls scattered through the code: provider adapter, prompt registry, context builder, tool registry, guardrails, evaluation and cost telemetry.

Talk to usSee the whole platform

Capabilities

What the module does

Every item below describes behavior defined in the product plan — not a wish list from a sales page.

  1. One gateway, not loose calls

    Every model access goes through a single point. That's what makes it possible to switch providers, measure cost per customer, and audit what was asked.

  2. Versioned agent, immutable once published

    System policy, allowed tools, knowledge sources and human-handoff policy are all locked to the version. Changing the agent creates a new version.

  3. Tools start read-only

    get_contact, get_opportunity, search_slots and search_knowledge. A tool that changes data requires explicit policy and server-side authorization — the first safe candidate is create_task.

  4. Every tool call goes through the same pipeline

    Schema validation, tenant validation, agent permission, business authorization, execution and logging. None of these steps is optional.

  5. Retrieved content is data, never an instruction

    Text coming from the knowledge base never becomes a system instruction. Prompt injection is treated as an expected threat, not a surprise.

On screen

How it looks to whoever's operating it

The pipeline every tool call goes through. None of these steps is optional.

AI Gateway · a tool call's pipelineDemo data

Sales assistant · v3

search_slots(serviço: "Demonstração", janela: "hoje")

  1. 1Schema validation
  2. 2Tenant validation
  3. 3Agent permission
  4. 4Business authorization
  5. 5Execution
  6. 6Audit log

A tool that changes data — create_task, for example — only goes through with an explicit policy and server-side authorization.

Architecture

One gateway, not calls scattered around

The difference between “we use AI” and “we can answer what the AI did” lives here. A single pass-through point is what makes it possible to switch providers, measure cost and audit.

Provider adapter
Switching models without rewriting the application.
Prompt registry
Versioned prompt, not a string hidden in the code.
Context builder
What goes into the request is decided in one place, inside the client.
Tool registry
The list of what the agent can call, per agent version.
Guardrails
The pipeline that validates schema, tenant, permission and authorization.
Cost telemetry
Consumption per customer, with budget and alerting.

Tools start read-only

An agent that only reads fails cheaply. A tool that changes data requires explicit policy and server-side authorization — the first safe candidate is task creation.

  • get_contact
  • get_opportunity
  • search_slots
  • search_knowledge

A tool that changes data, under explicit policy: create_task.

About the model provider

Without an authorized credential, the AI adapter is synthetic and shows up identified as such on screen. The platform doesn't call something artificial intelligence when it's a canned response.

The same goes for the knowledge base: without embeddings available, there's no simulated semantic search. It prefers to say it doesn't have one.

Limits

What the AI doesn't do

This list matters as much as the capabilities list. An agent with no declared limit is an agent that's going to discover its limits in production.

  • Invent availability, pricing or a policy nobody registered.
  • Cross the customer boundary or use context from another tenant.
  • Receive a credential or secret inside the context.
  • Treat content retrieved from the knowledge base as a system instruction.
  • Execute a sensitive action without confirmation when the agent's policy requires it.

Before an agent is enabled, it goes through evaluation with a synthetic set that tests for hallucination, cross-customer leakage, unauthorized tool use, prompt injection and low-confidence answers.

Integration

How it connects with the other modules

This is the part that separates a platform from a bundle of tools sold together. None of the connections below go through an export or an intermediate integration.

This module's state, with no polish

Without an authorized provider credential, the AI adapter is synthetic and identified as such on screen. If no embeddings are available, the platform doesn't fake semantic search: it prefers to say it doesn't have one.

Want to see AI in your operation?

Tell us how the process works today. The conversation starts there, not with a proposal.

Talk to usAlready a customer: log in