What Kvendra is built on

A few convictions shape every part of Kvendra. They explain why the core is open, why the agent never holds your secrets, and why your knowledge lives in a database you can read with plain SQL.

Open-core

The pieces that earn your trust — the knowledge-base engine, the skills, the broker — are open source, because trust and adoption come from being able to read and run the code, not from a promise. Opening the core is also how a tool spreads: developers reach for what they can inspect and fork. Value is captured in the things that are genuinely hard to replicate, not by locking up the parts you need to believe it works.

Encrypted local vault

The model orchestrating your work should not see a secret in the clear. In local mode, credentials you store in Kvendra stay in an encrypted local vault; when an operation needs one, the broker injects it at the moment of use and runs the operation on the agent's behalf, returning only the result. The boundary between the thing that reasons and the thing that holds the keys is enforced by the broker, not left to the model — the only way across it is the opt-in, audited unsafe.raw_token escape hatch. An operation that is not on the signed allowlist simply does not run.

A typed, server-side knowledge base

Scattered notes and free-text memory rot quietly. Kvendra's answer is a knowledge base where every fact is a typed entity, every change happens inside a transaction, and the engine writes a typed changelog server-side — so the history of why something is the way it is cannot drift out of sync with the data. That entity + transaction + changelog model is the defensible idea: structure your agent can query and reason over, with an audit trail it cannot forge.

Dogfooding

This very site is generated from a Kvendra knowledge base. The pages you are reading are produced by pulling typed entities and rendering them — so the documentation cannot quietly diverge from the system it describes. If Kvendra is good enough to build a product's developer portal from, it is good enough to manage the development behind it.

BYOK and MCP-native

Bring your own model. Kvendra does not lock you to a particular LLM; the agent that reasons is yours to choose. Everything speaks the Model Context Protocol — the Platform's tools and the broker's primitives — so your agent talks to Kvendra through an open, inspectable interface that is portable across the editors and agents that speak MCP.

Your data

Your knowledge lives in PostgreSQL with pgvector — a database you already know how to operate, back up and query. The whole core is self-hostable, so nothing has to leave your infrastructure when you use local embeddings, and a clean export means you are never locked in. The data is yours, in a place you control.

See how these come together in how it works, or read more in why Kvendra.