Llama CLI

A local AI assistant in your terminal. Chat with LLMs, attach files, run commands — all offline, all private.

Why

• Private: everything runs locally, no data leaves your machine
• Fast: no API latency, no rate limits, no subscriptions
• Integrated: read files, execute commands, and get AI-powered answers in one place

Install

Pre-built binary (recommended)

Download the latest release from GitHub Releases:

# Linux/macOS
curl -L https://github.com/ollama/llama-cli/releases/latest/download/llama-cli -o llama-cli
chmod +x llama-cli
./llama-cli

Or move it to your PATH:

sudo mv llama-cli /usr/local/bin/

From source

brew install ollama
brew services start ollama
ollama pull gemma4:e4b
make setup
sudo make install

Usage

hello                          # chat with the LLM
!ls src/                       # run command, output to terminal
!!cat src/main.cpp             # run command, output as LLM context
what does this code do?        # LLM can now see the file
/set                           # show runtime options
/set bofh                      # toggle BOFH spinner mode
/version                       # show version info
/help                          # show available commands
exit                           # quit

The LLM can also propose commands and file writes:

> fix the bug in main.cpp
[proposed: write src/main.cpp]
Write to src/main.cpp? [y/n/s]

The LLM can read files and make targeted edits:

> what does repl.cpp do?
[read src/repl/repl.cpp]          # LLM reads the requested file

> fix the typo on line 42
[proposed: str_replace src/repl/repl.cpp]
- old text
+ new text
Apply str_replace to src/repl/repl.cpp? [y/n]

Configuration

Setting	CLI arg	Env var	Default
Host	--host	OLLAMA_HOST	localhost
Port	--port	OLLAMA_PORT	11434
Model	--model	OLLAMA_MODEL	gemma4:e4b
Timeout	--timeout	OLLAMA_TIMEOUT	120
Exec timeout	--exec-timeout	LLAMA_EXEC_TIMEOUT	30
Max output	--max-output	LLAMA_MAX_OUTPUT	10000
No color	--no-color	NO_COLOR	auto-detect TTY
BOFH mode	--why-so-serious	—	false
System prompt	—	OLLAMA_SYSTEM_PROMPT	(built-in)

CLI flags override .env, .env overrides environment variables, which override defaults. Use --default-env > .env to generate a template configuration file.

See docs/user-guide.md for detailed configuration options.

Roadmap

• Connect to Ollama
• Configurable host, port, model, timeout
• Interactive chat with conversation memory
• Write files from response (<write>)
• Targeted edits (<str_replace>)
• Smart file reading (<read> with line ranges and search)
• Run commands (!, !!, <exec>)
• TUI: ANSI colors, markdown rendering, spinner
• Arrow key history (linenoise)
• Runtime options (/set markdown, /set color, /set bofh)
• Ctrl+C interrupt during LLM calls
• Stdin pipe support (--files)
• Auto-diff preview before file writes
• Streaming responses
• Inline code rendering in markdown

Quality Framework

This project follows ADR-048, a lean quality framework designed so that any AI agent — including small local models — can execute development tasks by following self-contained prompts with exact file paths, code, and verification commands.

The framework uses CMMI-inspired maturity levels (0-3) where every automated check serves at least two purposes. Current level: CMMI 0 (working toward CMMI 1). See ADR-048 §15 for the live audit.

Task prompts for AI agents live in docs/prompts/ — each prompt is copy-paste ready for a local Ollama model.

Contributing

See CONTRIBUTING.md for development setup and docs/ for architecture decisions.