The AI Agent That's Trying to Write Its Own Codebase
You've probably used AI to generate code snippets, debug errors, or even scaffold an entire project. But what happens when the AI-generated project itself becomes an AI agent, tasked with eventually maintaining its own repository? That's the experiment behind ANUS, a terminal-based AI agent that started as a prompt and is now aiming to become its own developer.
ANUS (Autonomous Networked Utility System) is a Grok-powered CLI agent built on an unusual foundation: its entire codebase was written by another AI, not a human. The project is a public experiment in what happens when you let software participate in its own evolution.
What It Does
ANUS is a command-line AI agent that runs in your terminal and works with your local files. It's powered by Grok, specifically optimized for the xai/grok-code-fast-1 model, and it ingests the context of whatever directory you launch it in. You start it with a simple anus command, and it reads your project structure to understand what you're working on.
The architecture is built around the Model Context Protocol (MCP), which gives it an extensible system for plugging in custom tools. Out of the box, it comes with integrated workspace tools for file system operations and sandboxed shell command execution. It also uses a local ANUS.md file to retain project-specific goals and instructions, so it can keep context across sessions.
Under the hood, it's a Node.js application requiring v20 or higher, installed globally via npm. You'll need an OpenRouter API key to use it, which you can obtain from their site.
Why It's Cool
Here's the thing that makes ANUS genuinely interesting: the project is a self-referential experiment. The README is upfront about this—the initial framework was generated by another AI agent named Manus, and from there, ANUS took on "a life of its own." That's not marketing spin; it's the core thesis of the project.
-
99% AI-generated codebase: The entire repository as it stands today was written by AI with limited human guidance. That's a bold claim and a fascinating baseline for any open-source project.
-
A three-phase roadmap toward self-development: The project isn't just building a static tool. Phase 1 is a capable Grok-powered CLI agent (current). Phase 2 introduces "meta-cognition," where ANUS analyzes its own codebase and suggests improvements by pulling in novel techniques from the broader AI ecosystem. Phase 3 is the ambitious endgame: ANUS becomes a core maintainer of its own repository, autonomously proposing, implementing, and shipping features with human oversight approaching zero.
-
It's a community-driven performance: The README frames this as a public exploration of AI development where every contribution is part of the experiment. You're not just using a tool; you're participating in a live test of whether an AI can eventually maintain itself.
-
Prompted, not programmed: The original system was created from a prompt rather than hand-crafted code. The goal is to return to that genesis by building an agent capable of contributing to its own evolution.
That last phase—autonomous self-maintenance—is either incredibly exciting or slightly terrifying depending on your disposition. Either way, it's a genuinely novel thing to watch unfold in public.
How to Try It
Getting started is straightforward if you have Node.js v20 or higher on your machine. Install the package globally:
npm install -g @anus-dev/anus
Then grab an API key from OpenRouter.ai. Once you have that, navigate to any project directory and run:
anus
The tool will walk you through API key setup on first launch, then automatically ingest the context of your current directory. From there, you can start issuing commands and working with your files through the agent.
If you want to follow the project's evolution or contribute to the experiment, the repository is at github.com/nikmcfly/ANUS. The README mentions contributions are welcome, and given the project's nature, your pull requests are literally part of the experiment.
Final Thoughts
ANUS is best suited for developers who are curious about the trajectory of AI-assisted development and don't mind being early adopters of an evolving tool. It's not trying to be the most polished agent on the market—it's trying to be the first agent that eventually doesn't need its human maintainers. Whether that succeeds or fails, it's a worthwhile experiment to witness (and maybe contribute to). The project is honest about where it is: a foundational agent with ambitious goals. Watching it try to climb that mountain is the real value here.
Follow @githubprojects for more developer tools and open source projects.