llm CLI Tool Guide 2026: Run Any LLM From Your Terminal

Installation & Quick Start

Install via PyPI. Python 3.8+ required.

pip install llm

Set your first API key:

llm keys set openai
# Paste your key at the prompt

Run your first prompt:

llm "explain vector embeddings in one sentence"

Every prompt is logged automatically. Review your history immediately:

llm logs

That's the full surface area of the core tool. Everything else—additional providers, local models, structured output—is additive through plugins.

Plugin System: Adding More Models

The plugin interface is where llm separates from every single-provider CLI. Each plugin is a standard PyPI package.

# Add Anthropic Claude
llm install llm-claude-3

# Add Google Gemini
llm install llm-gemini

# Add local models via Ollama
llm install llm-ollama

After installation, list every available model across all installed plugins:

llm models

Switch providers by passing the -m flag:

llm -m claude-3.5-sonnet "summarize this document"
llm -m gemini-1.5-pro "summarize this document"
llm -m ollama/llama3 "summarize this document"

Model aliases reduce friction further. Set a default model once:

llm aliases set default claude-3.5-sonnet

From that point, bare llm "prompt" calls route to Claude without any flag. Switching providers in a team context becomes a one-line alias update—not a codebase change. For full local model execution without any outbound API call, the Ollama integration documented at /tools/ollama is the reference path.

Prompt Logging & History

Every interaction—prompt, model, response, timestamp—writes to a local SQLite database. Location:

llm logs path
# Returns something like: /Users/yourname/Library/Application Support/io.datasette.llm/logs.db

Query recent logs:

llm logs
# Returns last 3 conversations by default

llm logs -n 20
# Last 20 entries

Export to JSON for programmatic processing:

llm logs --json

Because the store is SQLite, you can query it directly with any SQL client or with Datasette (Willison's own tool):

SELECT model, prompt, response, datetime(timestamp, 'localtime')
FROM responses
ORDER BY timestamp DESC
LIMIT 50;

Production use cases this enables include audit trails for LLM-assisted decisions, personal prompt pattern analysis, building a searchable knowledge base from AI-assisted research sessions, and cross-provider output comparison on identical prompts over time. The data never leaves your machine unless you explicitly export it.

Shell Scripting with llm

Three patterns surface repeatedly in scout log analysis of practitioner deployments:

Automated commit message generation

git diff --staged | llm "write a concise git commit message for this diff"

Add it as a git alias and it becomes a one-keystroke step in any commit workflow.

Error log diagnosis

tail -n 50 /var/log/app/error.log | llm "identify the root cause and suggest a fix"

Pipe the last 50 lines of any log file into llm. Works with any model—route complex traces to a high-context model, quick lookups to a faster or cheaper one.

Structured JSON extraction with jq

llm "extract the company name, ARR, and funding round from this text as JSON: $(cat deal_notes.txt)" \
  | jq '.company_name, .arr, .funding_round'

Combine llm's text output with jq to extract structured fields from unstructured input inside a pipeline. This pattern replaces a Python script with a one-liner in many data-wrangling contexts.

Frequently Asked Questions

How does llm compare to the OpenAI CLI?

The OpenAI CLI (openai) is provider-specific—it only calls OpenAI endpoints. llm is provider-agnostic by design: the same command syntax routes to any installed plugin backend. If your team is already using multiple providers or anticipates switching, llm's plugin architecture eliminates the need for separate CLI tools per vendor. The prompt logging to SQLite also has no equivalent in the OpenAI CLI.

Can I use llm with local models, with no data leaving my machine?

Yes. Install the llm-ollama plugin, run Ollama locally, and point llm at a local model with -m ollama/modelname. No API key required. No outbound network calls for the inference step. The prompt log writes to local SQLite only. The full local execution path is documented in HookFlow's Ollama tool profile.

What happens to my prompts—is anything sent to a server besides the model API?

Nothing beyond the model provider's API endpoint. The llm binary itself has no telemetry, no analytics calls, and no cloud sync. Your prompts go to whichever model backend you target (OpenAI, Anthropic, etc.) and are logged locally to SQLite. If you run a fully local model via Ollama, no prompt data leaves your machine at any point.

Track the Heat Score Live

The viral score and cross-platform signal data for llm—along with every other tool in the HookFlow dataset—updates continuously. If you're making a build-vs-buy decision on CLI tooling, provider abstraction layers, or local model infrastructure, the live score is the signal to watch.

→ Track the llm heat score and 40+ AI infrastructure tools live at hookflow.ai