terminus-ai 2.0.4

Terminus CLI: An autonomous AI agent for terminal-based task execution

Terminus CLI

Terminus CLI is a CLI agent for terminal-based task execution. It can be used as a standalone tool, although it was primarily designed as a research-preview agent for evaluating the abilities of language models to power autonomous agents in the terminal.

Note that Terminus CLI is a fork of Terminus-2 agent, which is a built-in agent in terminal-bench. Terminus CLI 2.0.0 shows slightly better performance than Terminus-2 agent with openai/gpt-5 model on terminal-bench@2.0.0 benchmark. Different from Terminus-2, which by design runs its control logic outside task's container, Terminus CLI is a standalone library that runs entirely inside task's container.

Installation

Prerequisites

• Python >=3.12
• tmux (required for terminal session management)

Installing tmux

Terminus requires tmux to manage terminal sessions. Install it using your system's package manager:

macOS:

brew install tmux

Ubuntu/Debian:

sudo apt-get install tmux

Fedora:

sudo dnf install tmux

Arch Linux:

sudo pacman -S tmux

Install Terminus

uv tool install terminus-ai

or

pip install terminus-ai

Usage

Command Line Interface

Terminus provides a CLI for quick testing and demonstration:

# Basic usage
terminus "Create a file hello.txt with Hello World"

# With options
terminus "Create a file hello.txt" \
  --model openai/gpt-4o \
  --logs-dir ./logs \
  --parser json \
  --temperature 0.7

# Show help
terminus --help

Note:

• The CLI runs directly on your local system using tmux (no Docker required)
• Perfect for quick tasks, testing, and automation

Programmatic Usage

You can also use Terminus programmatically in Python:

from terminus import Terminus
from pathlib import Path

agent = Terminus(
    logs_dir=Path("./logs"),
    model_name="anthropic/claude-sonnet-4",
    parser_name="json",  # or "xml"
    temperature=0.7,
    max_turns=100,
    enable_summarize=True,
)

Configuration Options

• model_name: The LLM model to use (required)
• parser_name: Response format - "json" or "xml" (default: "json")
• temperature: Sampling temperature (default: 0.7)
• max_turns: Maximum number of agent turns (default: 1000000)
• enable_summarize: Enable context summarization when limits are reached (default: True)
• api_base: Custom API base URL (optional)
• collect_rollout_details: Collect detailed token-level rollout data (default: False)