Skip to main content

LLM CLI

llm is a command-line tool for interacting with LLMs directly from the terminal. It is perfect for quick queries, piping data, and building shell-based workflows. If you work in the terminal (and as a data scientist, you should), llm is an essential tool.

Installation​

bash
# Install with uv
uv tool install llm

# Verify installation
llm --version

# Install a model plugin (e.g., for GPT-4o)
llm install llm-openai

# Set your API key
llm keys set openai
# Paste your key when prompted

Basic Usage​

bash
# Simple prompt
llm "What is the capital of France?"

# Use a specific model
llm -m gpt-4o "Explain quantum computing in one paragraph"

# Continue a conversation
llm -c "Now explain it even simpler"

# Pipe input
echo "def hello(): pass" | llm "Add type hints and a docstring to this function"

# Read from a file
llm "Summarize this article" < article.txt

# Save a conversation
llm "Explain Docker" --cid docker-explanation
llm -c "Give me a real-world example" --cid docker-explanation
Piping is powerful

The llm CLI works with Unix pipes, making it easy to integrate into existing workflows:

bash
# Summarize a log file
tail -100 app.log | llm "What errors are occurring?"

# Classify data
cat products.csv | llm "Categorize each product as 'electronics', 'clothing', or 'food'"

Prompt Templates​

Templates save reusable prompts with variable placeholders:

bash
# Create a template
llm templates edit code-review

# This opens your editor. Write:
# Review this code for bugs, style issues, and performance problems.
# Be specific and provide fixed code snippets.
#
# Code:
# {{code}}

Use the template:

bash
# With a file
llm -t code-review code="$(cat main.py)"

# With inline code
llm -t code-review code="def add(a, b): return a + b"

# List all templates
llm templates list

# Edit an existing template
llm templates edit code-review

Template with System Prompt​

bash
# Create a more complex template
llm templates edit sql-explain

# Template content:
# system: You are a SQL expert. Explain queries clearly with examples.
#
# Explain this SQL query in plain English, then suggest optimizations:
#
# {{query}}

Plugins​

llm has a rich plugin ecosystem:

bash
# List available plugins
llm plugins

# Install plugins
llm install llm-claude        # Anthropic Claude
llm install llm-gemini        # Google Gemini
llm install llm-mistral       # Mistral AI
llm install llm-ollama        # Local models via Ollama

# List available models
llm models

# Use a specific model
llm -m claude-3.5-sonnet "Write a haiku about data"
llm -m gemini-pro "What is 2+2?"

Logging and History​

Every interaction is logged to a SQLite database:

bash
# View recent conversations
llm logs

# View logs for a specific conversation
llm logs --cid docker-explanation

# Search logs
llm logs -q "Docker"

# Export logs as JSON
llm logs --json

# The database is at: ~/.llm/logs.db
# Query it with sqlite-utils:
sqlite-utils ~/.llm/logs.db "SELECT model, count(*) FROM responses GROUP BY model"

Batch Processing​

Process multiple items efficiently:

bash
# Process each line of a file
cat names.txt | while read name; do
    llm "Generate a professional email for: $name" >> emails.txt
done

# Or use xargs for parallel processing
cat items.txt | xargs -P4 -I{} llm "Classify: {}" >> classifications.txt

Embeddings​

llm also supports generating embeddings:

bash
# Install embedding model
llm install llm-openai
llm embeddings -m text-embedding-3-small "Hello world"

# Compare similarity between two texts
llm embeddings -m text-embedding-3-small "The cat sat on the mat" > vec1.json
llm embeddings -m text-embedding-3-small "A feline rested on a rug" > vec2.json
Combine llm with sqlite-utils

Since llm logs to SQLite, you can use sqlite-utils to analyze your LLM usage:

bash
sqlite-utils ~/.llm/logs.db "SELECT datetime(datetime) as date, model, substr(prompt, 1, 50) as prompt_preview FROM conversations ORDER BY datetime DESC LIMIT 10"