All Jina AI APIs as Unix commands. Search, read, embed, rerank - with pipes.
This CLI is designed for both humans and AI agents. An agent with shell access needs only run(command="jina search ...") instead of managing 20 separate tool definitions. The CLI supports pipes, chaining (&&, ||, ;), and --help for self-discovery.
pip install jina-cli
# or
uv pip install jina-cliSet your API key:
export JINA_API_KEY=your-key-here
# Get one at https://jina.ai/?sui=apikey| Command | Description |
|---|---|
jina read URL |
Extract clean markdown from web pages |
jina search QUERY |
Web search (also --arxiv, --ssrn, --images, --blog) |
jina embed TEXT |
Generate embeddings |
jina rerank QUERY |
Rerank documents from stdin by relevance |
jina classify TEXT |
Classify text into labels |
jina dedup |
Deduplicate text from stdin |
jina screenshot URL |
Capture screenshot of a URL |
jina bibtex QUERY |
Search BibTeX citations (DBLP + Semantic Scholar) |
jina expand QUERY |
Expand a query into related queries |
jina pdf URL |
Extract figures/tables/equations from PDFs |
jina datetime URL |
Guess publish/update date of a URL |
jina primer |
Context info (time, location, network) |
jina grep PATTERN |
Semantic grep (requires pip install jina-grep) |
The point of a CLI is composability. Every command reads from stdin and writes to stdout.
# Search and rerank
jina search "transformer models" | jina rerank "efficient inference"
# Read multiple URLs
cat urls.txt | jina read
# Search, deduplicate results
jina search "attention mechanism" | jina dedup
# Chain searches
jina expand "climate change" | head -1 | xargs -I {} jina search "{}"
# Get BibTeX for arXiv results
jina search --arxiv "BERT" --json | jq -r '.results[].title' | head -3jina read https://example.com
jina read https://example.com --links --images
echo "https://example.com" | jina readjina search "what is BERT"
jina search --arxiv "attention mechanism" -n 10
jina search --ssrn "corporate governance"
jina search --images "neural network diagram"
jina search --blog "embeddings"
jina search "AI news" --time d # past day
jina search "LLMs" --gl us --hl en # US, Englishjina embed "hello world"
jina embed "text1" "text2" "text3"
cat texts.txt | jina embed
jina embed "hello" --model jina-embeddings-v5-text-small --task retrieval.querycat docs.txt | jina rerank "machine learning"
jina search "AI" | jina rerank "embeddings" --top-n 5jina classify "I love this product" --labels positive,negative,neutral
echo "stock prices rose sharply" | jina classify --labels business,sports,tech
cat texts.txt | jina classify --labels cat1,cat2,cat3 --jsoncat items.txt | jina dedup
cat items.txt | jina dedup -k 10jina screenshot https://example.com # prints screenshot URL
jina screenshot https://example.com -o page.png # saves to file
jina screenshot https://example.com --full-page -o page.jpgjina bibtex "attention is all you need"
jina bibtex "transformer" --author Vaswani --year 2017jina pdf https://arxiv.org/pdf/2301.12345
jina pdf 2301.12345 # arXiv ID shorthand
jina pdf https://example.com/paper.pdf --type figure,tableEvery command supports --json for structured output, useful for piping to jq:
jina search "BERT" --json | jq '.results[0].url'
jina read https://example.com --json | jq '.data.content'| Code | Meaning |
|---|---|
| 0 | Success |
| 1 | User/input error (missing args, bad input, missing API key) |
| 2 | API/server error (network, timeout, server error) |
| 130 | Interrupted (Ctrl+C) |
Useful for scripting and agent workflows:
jina search "query" && echo "success" || echo "failed with $?"| Variable | Description |
|---|---|
JINA_API_KEY |
API key for Jina services (required for most commands) |
An agent with shell access can use this CLI directly:
result = run(command="jina search 'transformer architecture'")
result = run(command="jina read https://arxiv.org/abs/2301.12345")
result = run(command="jina search 'AI' | jina rerank 'embeddings'")No tool catalog needed. The agent discovers capabilities via jina --help and jina search --help. Errors include actionable guidance.
jina grep provides semantic search over files using local Jina embeddings on MLX. It requires a separate install:
pip install jina-grepjina grep "error handling" src/
jina grep -r --threshold 0.3 "database connection" .
grep -rn "error" src/ | jina grep "retry logic"Supports most GNU grep flags (-r, -n, -l, -c, -A/-B/-C, --include, --exclude) plus semantic flags (--threshold, --top-k, --model). Run jina grep --help for full options.
For repeated queries, start a persistent embedding server to avoid model reload:
jina grep serve start # background server, model stays in GPU memory
jina grep serve stop # stop when donejina embed, jina rerank, jina classify, and jina dedup support --local to run on Apple Silicon via the jina-grep embedding server instead of the Jina API. No API key needed.
# Start the local server first
jina grep serve start
# Local embeddings
jina embed --local "hello world"
cat texts.txt | jina embed --local --json
# Local reranking (cosine similarity on local embeddings)
cat docs.txt | jina rerank --local "machine learning"
# Local classification (cosine similarity on local embeddings)
jina classify --local "this is great" --labels positive,negative
# Local deduplication
cat items.txt | jina dedup --localLocal mode uses jina-embeddings-v5-nano by default. Override with --model jina-embeddings-v5-small.
Requires pip install jina-grep and jina grep serve start.
Inspired by CLI is All Agents Need:
- One tool, not twenty. A single
run(command="jina search ...")replaces a sprawling tool catalog. Less tool selection overhead, more problem solving. - Unix pipes are the composition model.
stdoutis data,stderris diagnostics. Commands chain with|,&&,||. No SDK needed. - Progressive
--helpfor self-discovery. Layer 0: command list. Layer 1: usage + examples. Layer 2: full options. The agent fetches only what it needs, saving context budget. - Error messages that course-correct. Every error says what went wrong and exactly how to fix it. One bad command should not cost more than one retry.
stderris the agent's most important channel. When a command fails,stderrcarries the fix. Never discard it. Never mix it with data.- Consistent output format. Same structure every time so the agent learns once, not every time.
--jsonfor structured, plain text for pipes. - Meaningful exit codes.
0success,1user error,2API error,130interrupted. Scripts and agents branch on these, not on parsing error strings. - Layer 1 is raw Unix, Layer 2 is for LLM cognition. Pipe internals stay pure (no metadata, no truncation). Formatting and context only at the final output boundary.
Apache-2.0
