AWorld CLI
AWorld
AWorld's terminal AI agent: an orchestrator runs a developer, an evaluator and media generators in a build, evaluate, evolve loop
- 1,244 ★ on the repository
- Last commit
- MIT
Broad access
Why this level
- The agent writes and runs code and works with files and a browser in your working folder
- Model and generation service keys live in `.env` or the config, and sub-agent jobs spend paid tokens and money on external services
- The self-improvement loop can rerun generation many times and burn limits quickly
Install
In your terminal, with SkillFoxx CLI
npx skillfoxx add cli/aworldDetects the agents on your machine, shows the access level and pins the version.
Other ways to install
Access and checks
How we reviewBroad accessThe tool can change data, run commands or handle secrets.
- The agent writes and runs code and works with files and a browser in your working folder
- Model and generation service keys live in `.env` or the config, and sub-agent jobs spend paid tokens and money on external services
- The self-improvement loop can rerun generation many times and burn limits quickly
This is third-party code. Review the repository files before installing.
What it does
AWorld CLI starts in the terminal with `aworld-cli` and runs a task through a team of sub-agents. The main AWorld agent breaks down the goal and delegates: Developer writes, debugs and refactors code, Evaluator checks the output against set criteria, and Video Diffusion, Audio Generator and Image Generator produce video, speech and images from text. The inclusionAI project also serves as a framework for building your own agents; this entry covers the terminal agent. The main scenario is a build, evaluate, evolve loop. You ask for, say, an English word-learning mini-app with a UI quality score above 0.9. Developer writes the HTML, Evaluator checks it with its app_evaluator skill, and if the score is below target Developer gets the list of issues and edits the code until the criterion is met. The agent's know-how is packaged as skills in the `aworld-skills` folder: app_evaluator, agent-browser, filex, embedded_video, text2agent, self_evolve, long-running-agent, workbench and more. For large codebases the project uses CAST (Code Abstract Syntax Tree): the agent sees a blueprint of the code and gets only the relevant parts. Built-in CLI plugins live in `aworld-cli/src/aworld_cli/builtin_plugins`.
Who it is for. For developers who want a terminal agent with a self-check loop and ready skills and are ready to connect their own model.
Good fit when
- You want to build a small web app and drive it automatically to a target UI quality score
- The task chains code, a result check, then video, speech or images
- You want to write and plug in your own skills for a terminal agent and watch how it evaluates results
- You want to experiment with agents that improve other agents through the text2agent and self_evolve skills
Not a fit when
- You want a skill or plugin for Claude Code or Cursor you already use: AWorld CLI runs its own agent and does not embed in them
- You have no keys for a model and generation services: without them the agent and the Video, Audio and Image sub-agents do not work
- You want a one-command install without cloning: the CLI installs only from source
- The task is simple and one model call solves it: a multi-agent setup with an evaluator wastes tokens
Example request
Build a mini-app for learning English words and drive the UI quality score to 0.9Limitations
- Install is from a repository clone: `pip install -e .` at the root, then `pip install -e .` in `aworld-cli`. The authors suggest a conda environment with Python 3.11.
- The model and API settings come from `aworld-cli --config` or a `.env` file in the working folder.
- The video, audio and image sub-agents use external generation models (for example Kling-V3 for video), which need their own keys and paid balance.
- The project is large and includes a framework, benchmarks and the CLI. The authors call the documents under `aworld-cli/internal/docs/superpowers/` historical planning notes.
How to disable. Remove the packages with `pip uninstall aworld-cli` and `pip uninstall aworld` in the same environment, and delete the repository clone and the `.env` file with keys by hand.
FAQ
How do I install AWorld CLI?
Clone inclusionAI/AWorld, create a Python 3.11 environment and run `pip install -e .` at the root, then `pip install -e .` in `aworld-cli`. `aworld-cli --config` sets up the model.
Does AWorld CLI plug into Claude Code?
No, it is a standalone terminal agent with its own sub-agents and the skills in `aworld-skills`. The skills use the SKILL.md format, but the project gives no ready way to install them into Claude Code.
Which sub-agents does AWorld CLI have?
The main AWorld Agent, Developer, Evaluator, Video Diffusion, Audio Generator and Image Generator. The first three handle code and checks, the others generate media.
What is the build, evaluate, evolve loop?
Developer writes a solution, Evaluator scores it with the app_evaluator skill, and if the score is below target Developer gets the notes and edits the code. The loop repeats until the criterion is met.
How do I set the model and keys?
Run `aworld-cli --config` or create a `.env` file in the working folder with the model and API variables. The core variables are listed in `docs/AWorld CLI/Configuration.md`.
Discussion
No questions or reviews yet.
Related
CLI toolsAn open coding agent for the terminal, browser and desktop that works with models from OpenAI and Claude to Ollama
A team of AI agents from one request: it plans the task, writes code in parallel branches, tests, reviews and opens a pull request
A CLI installer and package manager for coding agent skills: one command installs skills into many runtimes and helps build your own
A LangWatch command-line tool that creates the scaffold of a new AI agent project and hands it to a coding assistant
A skill with 25 commands to audit and improve the prompts, tools and agent designs of a project in Cursor, Claude Code and Codex
A CLI to run several coding agents: one config, a session console, local memory and a set of engineering skills
Comparison with similar tools
Cards that share the main search query, and how they differ.
| Tool | Official | Repository stars | Main difference |
|---|---|---|---|
| AWorld CLI (this card) | no | 1,244 | |
| goose | no | 56,269 | an open Rust agent with a desktop app and MCP extensions, while this one centers on a result-evaluation loop with an evaluator skill |
| pydantic-deepagents | no | 1,115 | a terminal assistant with a Docker sandbox, while this has a sub-agent team with media generation |
README in short
The README shows works that the AWorld agent built from skills: apps, video and search. It then explains CLI install, configuration and sub-agent roles, the self-evolution loop, the CAST technique and the shared skill system that the community adds expertise to. Sections on the framework, benchmarks and the roadmap follow.