kirya.
Under the hood

How it works, under the hood.

The technical detail of kirya for evaluators: what it connects to, what it reads, how it is organised inside and where it runs.

What kirya connects to

GitHub, Azure, PostgreSQL, Google Workspace and Asana, with incremental sync. Each connector brings observations from the source system into the graph. New connectors ship as packages, without waiting for a new platform release, and the list grows with customers' environments.

What kirya reads. And what it doesn't.

To build the map, kirya reads metadata: schemas and column names, repositories and their owners, configuration files, cloud resources, projects, declared documents. It does not read production rows. It does not ask for write access. Sensitive data is classified by structure, with no copy of content.

AI models

kirya does not sell tokens. You configure your own key (Azure OpenAI, OpenAI, Anthropic or compatible) and billing stays on your account, with no intermediation. Initial discovery runs on that key and costs metadata, not content. Routing by task type and per-area budgets are rolling out.

Where it runs

Multi-tenant SaaS, with per-tenant isolation on every query. Single-tenant or on-premise for regulated environments. Local hosting and local invoicing available in Brazil.

The five layers

  1. ConnectorsObservations from the source systems.
  2. Technical ontologyA canonical graph with a closed vocabulary: eight node types (person, code, work, application, infrastructure, data, knowledge, tool) and five relations (contains, deployed to, has access to, depends on, references).
  3. Domains and semantic layerBusiness zoning over the graph. What the AI infers enters as pending and only becomes confirmed by a responsible person's decision. Each company teaches its own vocabulary: synonyms, nicknames, product names. Guidelines, org charts and documents enter as versioned evidence, never as presumed fact.
  4. Policies and permissionsDefault deny: what is not mapped is not exposed to the model. Three roles (read, curation, administration), scoped grants per node or domain, sensitive-data masking, an audit trail on every call and every state change.
  5. ConsumptionAn MCP endpoint your team's tools consume: Cursor, Claude Code, Lovable, Copilot. The kirya interface for non-technical users. The CTO console.

Want the full technical detail for your security team? Request the architecture document