Documentation · Documentation
For AI agents (MCP)
Give Claude Code, Cursor, Claude Desktop or VS Code local tools for folder search, transcription, classification, reranking and similarity with kjarni mcp.
kjarni mcp serves Kjarni to any MCP client as tools that run on your machine, with no
API key. It comes with the command line tool.
Add it to your client
Claude Code:
claude mcp add kjarni -- kjarni mcp
Cursor, Claude Desktop, VS Code and other clients take the same command in their MCP configuration:
{ "mcpServers": { "kjarni": { "command": "kjarni", "args": ["mcp"] } } }
Tools
| Tool | What it does | Arguments |
|---|---|---|
search_folder | Finds passages in a folder of documents or code, by meaning and by keyword | path, query, mode (hybrid, semantic or keyword), top_k, rerank |
transcribe | Speech to text with Whisper, with timestamps. Takes wav, mp3, flac or ogg | path |
classify | Sentiment, emotion or toxicity of a text | text, model |
rerank | Orders passages by how well each answers a query | query, documents |
similarity | Cosine similarity of two texts, from -1 to 1 | a, b |
path is an absolute path. Models download the first time a tool needs them and stay
loaded for the session. A first download can take longer than a client's tool timeout,
so fetch large models beforehand:
kjarni model download minilm-l6-v2
kjarni model download whisper-small
Folder search
The first search of a folder builds an index, and later searches reuse it until the
files change. Indexes are kept in Kjarni's cache, never in your folder. From 0.2.8,
build output, dependencies and hidden folders are skipped: target, node_modules,
dist, build, out, bin, obj, vendor, venv, __pycache__, *.min.js and
anything starting with a dot.
Privacy
The models run locally, and nothing is sent anywhere by Kjarni. The results do go to
your agent, so a passage from search_folder or a transcript reaches whatever model
that agent uses. If the agent runs in the cloud, treat tool results as you would any
file you show it.
Kjarni runs the same engine on every platform: