A chibi-sized summary for any text. Pipe text in, get a structured summary out. This tool will only do the summarization. For turning your file into text use other tools like pdftotext, kreuzberg, typeout, ...
$ go install github.com/miku/chibi/cmd/chibi@latest
Or packaged as deb or rpm.
If no explicit LLM endpoint and model are given, chibi will try to discover
a suitable endpoint and model by looking for typical environment variables,
like OPENAI_BASE_URL or OLLAMA_HOST, but you can also set endpoint -e
and model -m explicitly.
To show the discovered endpoint and model:
$ chibi -i
endpoint: http://chiba:11434/v1
model: nemotron-3-nano:30b-a3b-fp16
styles: article, basic, paper, paperplus, podcast, transcript
cache: /home/tir/.cache/chibi
$ kreuzberg extract testdata/2025.loreslm-1.13.pdf | \
chibi -s article -m qwen3.6:latest | glow
$ typeout https://www.youtube.com/watch?v=S4EsRyZQKEc | \
chibi -s transcript -m qwen3.6:latest | glow
See some example rendering/screenshot below.
# Default summary using the autodiscovered model
$ cat article.md | chibi
# Pick a style tailored to the input type
$ chibi -s transcript < lecture.vtt
$ chibi -s podcast < interview.txt
$ chibi -s paper < paper.txt
# Like paper, but weaves in short verbatim quotes from the source
$ chibi -s paperplus < paper.txt
# Show resolved endpoint, model, styles, cache dir (no LLM call)
$ chibi -i
# Force a fresh variant (LLMs are probabilistic; -f grows the cache)
$ chibi -f < input.txt
# Render every cached variant for an input as one markdown document
$ chibi -A < input.txt | glow -p -
# Use a specific model on a remote endpoint
$ chibi -e https://api.example.com/v1 -k "$TOKEN" -m gpt-4o < article.txt
Using various tools to turn files into plain text.
