How it works

How the pieces fit together

Kvendra is three open-core pieces wired together over the Model Context Protocol (MCP), with one in the middle: your AI coding agent. The core is Kvendra Platform — a self-hosted knowledge base. Skills run on top of it, and an optional CLI adds a security broker. Everything below is something you can read, run and own.

The three pieces

Kvendra Platform is the knowledge base and the only required piece. It runs as an MCP HTTP server on port 7777, backed by PostgreSQL with the pgvector extension. It holds typed entities, runs transactions, and writes a server-side changelog, so the things your agent learns about a project live in a database you control — not in scattered notes.

Skills are a Claude Code plugin. They turn pipelines — planning a feature, running tests, cutting a release — into one-command workflows by calling the Platform's tools. The recipes live in the knowledge base, so the skills stay thin and portable.

CLI & Broker is the optional, pre-alpha security layer. It runs a second MCP server over stdio: a zero-knowledge vault plus a capability broker that performs credentialed operations on your behalf, so the model never sees a raw token. Skills and the Platform work without it; you add it only when a workflow needs brokered system operations.

How they are wired

Your agent (Claude Code, the recommended host) speaks MCP. It talks to the Platform to read and write the knowledge base, and — if it is installed — to the CLI broker to run secured operations. The diagram below traces the wiring.

Core · required Kvendra Platform MCP HTTP · :7777 Postgres + pgvector · typed entities · TXN · changelog

Skills Claude Code plugin · pipeline workflows

Claude Code the host · speaks MCP

Optional · pre-alpha CLI & Broker MCP stdio zero-knowledge vault · capability broker

Claude Code reads and writes the knowledge base on the Platform over MCP HTTP, and (when present) routes secured operations through the CLI broker over MCP stdio.

The flow, step by step

Your agent needs project context, so it queries the Platform over MCP — semantic search and typed lookups across the knowledge base.
As work lands, the agent opens a transaction and creates or updates entities; the Platform records every change in a typed changelog.
A skill orchestrates the larger workflow — a feature pipeline, a release — by chaining those same Platform tools.
When a step needs a credentialed operation (push to a remote, call an API), and the optional CLI is installed, the agent asks the broker to run it. The broker resolves the secret from the vault and never hands it back.

Ready to wire it up? The Quickstart walks through it in order — Platform first, then the rest.