Command line

kjarni runs transformer models from your shell: classify text, search a folder by meaning, find photos by description, transcribe audio, summarize, translate and chat. Everything runs on your machine. Each model downloads once, the first time a command needs it.

Install

curl -fsSL https://kjarni.ai/install.sh | sh

On Windows, in PowerShell:

irm https://kjarni.ai/install.ps1 | iex

The installer puts kjarni in ~/.local/bin and adds that folder to your PATH, so a new terminal finds it. Linux (x86-64 and ARM), macOS on Apple Silicon and Windows are supported. More options are on the install page.

kjarni --version
kjarni --help             # every command
kjarni classify --help    # every option of one command

The first run downloads the model

Every command uses a default model, and you can pick another with -m. The first time a model is needed, kjarni downloads it and shows the progress:

  Repository: sentence-transformers/all-MiniLM-L6-v2
  Size:       ~90 MB
    model.safetensors 74% (0.1 / 0.1 GB)
✓ Download complete!

After that it loads from ~/.cache/kjarni and works offline. The models page lists every model by name, and kjarni model list shows which ones you have.

Classify text

Sentiment by default:

kjarni classify "I love this product"

  Input "I love this product"

  ✓       POSITIVE  ████████████████████  100.0%
          NEGATIVE  ░░░░░░░░░░░░░░░░░░░░    0.0%

Pick another classifier with -m, such as roberta-emotions for 28 emotions:

kjarni classify -m roberta-emotions "I can't believe I won!"

  ✓       surprise  ████████████████░░░░   78.3%
        excitement  █████░░░░░░░░░░░░░░░   26.4%
           neutral  █░░░░░░░░░░░░░░░░░░░    4.1%
               joy  █░░░░░░░░░░░░░░░░░░░    3.9%
       realization  ░░░░░░░░░░░░░░░░░░░░    2.5%

Several words without quotes are one text. To classify several texts, give one per line on stdin:

printf 'I love this product\nThe box arrived crushed\n' | kjarni classify

        POSITIVE  ███████████████  100.0%  "I love this product"
        NEGATIVE  ███████████████  100.0%  "The box arrived crushed"

--format json or --format jsonl gives output a script can read. Other classifiers: roberta-sentiment (negative, neutral, positive), bert-sentiment-multilingual (1 to 5 stars, six languages) and toxic-bert. kjarni model search sentiment finds them by name.

Compare two texts

kjarni similarity "a cat sat on the mat" "a kitten rested on the rug"

  ████████████░░░░░░░░   61.8%  moderately similar

  ↔ "a cat sat on the mat"
  ↔ "a kitten rested on the rug"

Embeddings

kjarni embed prints the vector, normalized to unit length:

kjarni embed "hello world"

-0.03447726 0.031023178 0.0067349467 0.026109012 -0.039362032 -0.16030246 ...

--format json prints a JSON array. The default model, minilm-l6-v2, gives 384 numbers, the same as sentence-transformers. Pick another with -m, such as nomic-embed-text or the multilingual bge-m3.

Rerank passages

Score candidate passages against a question with a cross-encoder. The output is score, original position and text, best first:

kjarni rerank "how do I reset my password" \
  "Click Forgot password on the sign-in page." \
  "Our office is closed on Sundays." \
  "Passwords must be 12 characters long."

-3.6341	0	Click Forgot password on the sign-in page.
-7.5938	2	Passwords must be 12 characters long.
-11.2764	1	Our office is closed on Sundays.

Higher is more relevant. With no passages on the command line, it reads one per line from stdin. -k 5 keeps the best five.

Search a folder

Build an index of a folder once, then search it by meaning:

kjarni index create notes.idx notes/

✓ Index created: notes.idx
  Documents: 3
  Dimension: 384
  Size: 9.65 KB

kjarni search notes.idx "the heating is losing pressure" -k 2

  Results for "the heating is losing pressure"

    1. notes/boiler.md
       ████████████████████  100.0%
       "# Boiler service The boiler is serviced every October. Call Hamar Heati…"

    2. notes/wifi.md
       ████████████████████   98.4%
       "# Home network The router is in the hall cupboard. Guest network: Hus-G…"

The note never says "heating" or "losing pressure"; it matched on meaning. The bar is relative to the best result, which is always 100%.

--mode chooses how results are ranked:

• hybrid, the default, combines meaning and keywords, which is usually best.
• semantic uses meaning only, and its bars show the gap better: here the Wi-Fi note drops to 26%.
• keyword uses exact words only (BM25).

Add --rerank-model minilm-l6-v2-cross-encoder to re-score the top results with a cross-encoder, which is slower and more precise. --format json includes each result's file path and chunk.

To keep an index current:

kjarni index add notes.idx notes/tyres.md       # add a file, or replace it if it changed
kjarni index remove notes.idx notes/tyres.md    # drop a file
kjarni index info notes.idx                     # documents, size, a sample

Text, Markdown, HTML, JSON, CSV and source code files are indexed; PDF and Word files are not read yet. Long files are split into chunks of 512 characters, which --chunk-size changes.

Find photos by description

kjarni image index photos/ -o photos.json
kjarni image search photos.json "snow on a mountain" -k 2

  Results for snow on a mountain

    1. photos/snowy_mountain.png
       ████████████████████  0.2005
    2. photos/deep_forest.png
       ███████████████░░░░░  0.1535

The search uses CLIP and reads the pictures themselves, not their file names. It reads JPEG and PNG, in subfolders too, but not iPhone HEIC photos. Indexing takes about 20 ms a photo on a desktop CPU, and every search after that reuses the index.

Transcribe audio

kjarni transcribe kennedy.wav

 And so my fellow Americans, ask not what your country can do for you, ask what you can do for your country.

WAV, MP3, FLAC and Ogg Vorbis files work, at any sample rate. The default is whisper-small, and -m whisper-large-v3 is more accurate, especially outside English.

• --timestamps prints each segment with its time: [00:00.000 --> 00:11.000] And so my fellow Americans, ...
• -l is sets the language instead of detecting it.
• --translate gives English text from speech in another language.

Summarize

kjarni summarize -i "The city council voted on Tuesday to extend the harbour ferry service through the winter. Ferries will now run every two hours between October and March, instead of stopping entirely. ..."

 Ferries will now run every two hours between October and March, instead of stopping entirely . Council said demand from commuters had doubled over the past three years . ...

With no -i, it reads stdin, so kjarni summarize < article.txt works. The default model is distilbart-cnn; bart-large-cnn is larger and better. Both were trained on news articles.

Translate

kjarni translate --src en --dst de -i "The ferry leaves at nine."

Die Fähre fährt um neun Uhr.

English to German and German to English use dedicated Opus-MT models and are reliable. Other pairs fall back to flan-t5-base, which is much weaker.

Generate text and chat

kjarni generate --gguf "Three tips for a long flight:"

 Stay hydrated, avoid screens before landing, and get some fresh air

--gguf uses the smaller quantized download, which is also much faster: that answer took about three seconds on a desktop CPU, loading included. The default model is llama3.2-1b-instruct. Others include qwen2.5-1.5b-instruct, llama3.2-3b-instruct, phi3.5-mini-instruct and mistral-7b-instruct.

• -n 200 sets the most tokens to generate.
• -t 0.2 sets the temperature, and --greedy makes the output repeatable.
• --gpu runs on the graphics card through Vulkan, DirectX 12 or Metal, so AMD, Intel and Apple GPUs work as well as NVIDIA.

For a conversation that remembers earlier turns:

kjarni chat --gguf -m llama3.2-1b-instruct

Kjarni Chat: llama3.2-1b-instruct
Type '/help' for commands, '/quit' to exit.

> What is the capital of Iceland? One sentence.
Reykjavik is the capital and largest city of Iceland.

-s "You are a terse assistant" sets a system prompt.

Use it in scripts

Results go to stdout and progress goes to stderr, so pipes and redirects get only the output:

kjarni embed "hello world" > vector.txt
kjarni transcribe meeting.mp3 > meeting.txt
cat reviews.txt | kjarni classify --format jsonl > scores.jsonl
kjarni search notes.idx "boiler" --format json | jq -r '.[0].metadata.source'

-q turns off the progress messages entirely. A command that fails exits with a non-zero status and says why on stderr.

Use it from an AI agent

kjarni mcp gives Claude Code, Cursor and other MCP clients these tools: folder search, photo search, transcription, classification, reranking and similarity. See For AI agents (MCP).

Manage models

kjarni model list                  # every model, and which you have
kjarni model search emotion        # find one by name or description
kjarni model info minilm-l6-v2     # size, format, where it is stored
kjarni model download whisper-large-v3
kjarni model remove whisper-large-v3

Models are stored in ~/.cache/kjarni. To use them on a machine with no internet, see offline use.