Block Unveils 'Goose': New Open-Source AI Agent Allows Mid-Air Direction Changes

The wave of artificial intelligence (AI) seems to be cresting in favor of open-source solutions, and that's great news for both developers and everyday users. On Tuesday, Block, the company founded by Jack Dorsey behind Square, Cash App, Tidal, and Afterpay, unveiled a complete overhaul of its open-source AI developer agent, codename goose. Previously available in beta, this new version is released under the Apache License 2.0 (ASL2), promising to bridge the gap between user interfaces, language models, and systems. Thanks to Anthropic's ModelContext Protocol (MCP), goose can seamlessly connect to various systems, allowing it to tap into data stores, developer environments, and business applications. What's more, users have the freedom to choose any large language model (LLM) they wish to integrate with goose.

Wondering what this means for you? Well, for starters, the framework is open for both individual and commercial use, and it's incredibly flexible. Unlike closed-source tools that require you to specify all tools at the start, goose lets you bring in new tools mid-session. Imagine being deep into a conversation and suddenly realizing you need to pull data from Google Docs or connect to Slack. With goose, you can simply click through, enable a new extension, and voila! Goose is now equipped with these new capabilities, ready to run with them.

According to lead engineer Brad Axen in an interview with ZDNET, "You can be in the middle of a conversation, realize that you need to pull in something from your Google Docs or connect it to Slack or something, and you can just, say, click through, enable a new extension, and then goose suddenly has these new capabilities it can go run with." This level of adaptability is a game-changer, especially for developers who are constantly on the move.

Designed primarily as an engineering agent, goose is akin to platforms like Blitzy, excelling at real-time code searching and writing while autonomously executing tasks. From reading and writing files to running code and tests, refining outputs, installing dependencies, and handling additional actions, goose does it all, as the release claims.

To enhance its embeddability, the team switched goose's language from Python to Rust and crafted a native application, with goose itself contributing 70% of the code. Axen highlighted that codename goose "saves 20% of [developers'] time, or they'll find tasks where it can do a day or two worth of work all by itself, unsupervised."

But Block's vision for goose extends beyond just developers. Jackie Brosamer, head of AI platform at Block, shared that they aim for goose to be versatile enough for creative pursuits, like music generation. "We really built a lot of the fundamental use features as kind of like a general purpose capability," Brosamer explained. "A big part of the way we're able to expand beyond the developer use cases is our open source philosophy."

To underscore its commitment to open-source, Block launched an Open Source Program Office last week. Brosamer pointed out that integrating with Anthropic's MCP allows users to connect goose with everyday apps like email, Google Calendar, and Drive. "We're starting to see people use codename goose to do things like simplify your day and try to get summaries of everything you read next last week," she added. Within Block's financial ecosystem, she envisions customers using goose to automate tasks like splitting bills through Cash App, managing monthly business accounts, or even ordering dinner via Square.

Brosamer also noted that goose is empowering less technical users. "We're already seeing designers able to prototype their own front-ends without having to bring in an engineer."

Looking ahead, Block sees goose evolving into a daily AI assistant. "I could see a place where, as we get smaller models that can do the agent loop effectively, we start to move those on device," Axen said. "It's pretty easy for us to take the agent path of goose, accompany it with an on-device model, and put it onto an Android phone or something like that."

The release of codename goose follows closely on the heels of DeepSeek's R1, an open-source model rivaling OpenAI's o1 at a fraction of the cost, and AI assistants from Perplexity and OpenAI. Goose is well-positioned to meet both commercial and personal needs, harnessing the vast potential of open-source.

"If someone has a good idea, they don't have to wait on us to build it," Axen remarked. Brosamer added, "Anyone can contribute new capabilities to the agent itself."

You can dive into codename goose and explore all its documentation here.