AI Agents (MCP)

kura is designed to be an AI agent’s long-term memory. It speaks the Model Context Protocol over stdio, so an agent can search your knowledge base and add to it as part of its normal tool use.

Connect it to Claude Code

claude mcp add kura -- kura mcp

For any other MCP client, print a ready-made config snippet:

kura mcp --print-config     # prints an .mcp.json entry

Then run kura mcp — it serves on stdio and stays local; no network, no account.

Exposed tools

Tool	Purpose
kura_query	Hybrid search (keyword + semantic + rerank) — the default retrieval tool
kura_search	Keyword search (FTS5 BM25)
kura_get	Fetch a document by key, full path (clips/Title), or title
kura_add	Create a document; an optional path files it under a folder-like document path
kura_update	Update an existing document; changing title or path rewrites [[links]] in referring documents
kura_list_tags	List the tag hierarchy
kura_related	Documents related to a given one (links + similarity)
kura_status	Store statistics

An agent typically reaches for kura_query to recall context, kura_related to explore around a hit, and kura_add / kura_update to write findings back — turning kura into durable memory across sessions.

Degraded operation

The MCP tools follow the same rule as the CLI: kura_search works with no LLM provider, and kura_query falls back toward keyword results when semantic search or reranking is unavailable (see Search). An agent never gets a hard failure just because no model is running.

--json everywhere

If you’d rather script kura than speak MCP, every read command supports --json:

kura query "SQLite concurrency" --json
kura get "Today's note" --json
kura status --json

The JSON shapes are stable contracts — the same payloads back the MCP tools — so you can build on them safely.