Goose is an open-source, model-agnostic AI agent for software engineering created by Block (formerly Square). It runs in your terminal and can install packages, execute commands, edit files, and run test suites autonomously — going beyond suggestions to actually perform engineering tasks end-to-end.

How does Goose differ from GitHub Copilot or Claude Code?

Copilot and Claude Code provide in-editor suggestions and limited terminal actions. Goose is a fully autonomous command-line agent that can take multi-step actions: it reads your repo, plans a solution, writes and executes code, installs dependencies, runs tests, and iterates on failures — all without human confirmation at each step.

Which AI models does Goose support?

Goose is model-agnostic. You can use it with Claude (Anthropic API), GPT-5.4 (OpenAI), Gemini 3 Pro (Google), and any OpenAI-compatible API. You bring your own API key — Goose does not lock you to a specific provider or charge a subscription on top of your LLM costs.

Is Goose safe to run on production codebases?

Goose is designed for development environments, not production systems. It executes commands directly on your machine, so Block recommends using it in isolated dev containers or sandboxed environments, especially for unfamiliar codebases. The project is MIT-licensed and under active development — test thoroughly before using it in CI/CD pipelines.

Block Launches Goose: The Open-Source AI Agent That Actually Runs Your Code

Block open-sourced Goose on April 7, 2026 — a model-agnostic AI agent that installs, executes, edits, and tests code directly in your terminal. Unlike Copilot or Claude Code, it runs locally with your own LLM key and no cloud dependency.

What Goose Actually Does

Most AI coding tools are glorified autocomplete. You accept a suggestion, run the code yourself, read the error, paste it back, and ask again. Goose breaks that loop.

Goose is an autonomous agent. You give it a task — "add rate limiting to the API" or "fix all TypeScript errors in this repo" — and it plans the steps, executes them in your terminal, and iterates until the task is done or it hits a blocker it cannot resolve alone.

Concretely, Goose can:

Read any file in your repository
Execute shell commands (including npm install, pytest, cargo build)
Write and overwrite files
Run test suites and read failure output
Iterate on failing tests autonomously until they pass
Search the web for documentation when needed

This is closer to how a human developer works than any code assistant has been before.

Block Built This for Its Own Engineers

Block (formerly Square) built Goose for internal use first. The company employs thousands of engineers across Cash App, Square, Tidal, and TBD (its Bitcoin and Web5 division). The complexity of maintaining services across those products demanded something more capable than an IDE plugin.

The decision to open-source Goose follows Block CEO Jack Dorsey's stated philosophy on AI and organizational structure — he has advocated for fewer middle layers and more direct execution, which Goose embodies in software form.

The agent is now maintained publicly at github.com/block/goose with active PRs from the community from day one.

Model-Agnostic by Design

The most strategic decision Block made was to not build Goose around a single model. It supports any provider via OpenAI-compatible APIs:

Claude Opus 4.6 — best for multi-step reasoning on complex codebases
GPT-5.4 Thinking — strong for test generation and iterative debugging
Gemini 3 Pro — 1M context window for large monorepos
Llama 4 / Gemma 4 — run locally with no API cost at all

You configure your provider in a YAML file and Goose routes every task through it. Switching models is a one-line change.

Goose vs. Competing AI Coding Agents

Tool	Autonomous Execution	Model Lock-in	Open Source	Price
Goose (Block)	Full terminal agent	None — any LLM	Yes (MIT)	Free + LLM API costs
Claude Code	Full terminal agent	Claude only	No	$17–$200/mo or API
GitHub Copilot	Editor + limited agent	Partial (multi-model)	No	$10/mo
Cursor	Editor + agent mode	Partial (multi-model)	No	$20/mo
OpenHands	Full agent	None — any LLM	Yes (MIT)	Free + LLM API costs

How to Get Started in 5 Minutes

Goose requires Python 3.10+ or can be installed via pipx. The simplest path:

# Install
pipx install goose-ai

# Configure your LLM (example: Claude)
goose configure

# Run your first task
goose run "Write unit tests for all functions in src/utils.py"

Goose will read your file, generate tests, write them to a new test file, run them, and iterate on any failures — all without further prompting.

The Plugin System

Goose ships with a toolkit extension system. Core plugins available at launch:

Developer toolkit — file I/O, shell, grep, test runner
Web search — pulls documentation on the fly
GitHub — reads issues, creates PRs
Memory — persists context across sessions so it remembers your coding conventions

Third-party plugins can be registered via a simple manifest file, and Block has already seeded a community repository of plugins for popular stacks (Django, Rails, Next.js, FastAPI).

Who This Is For

Goose targets three groups:

Solo developers and solopreneurs who want a free agent they can run with their own API key and no vendor lock-in
Open-source contributors who want to automate repetitive maintenance tasks (dependency updates, test coverage, linting fixes)
Engineering teams exploring autonomous coding infrastructure before committing to proprietary agents

If you are already paying for Happycapy or Claude Code, Goose is a complementary tool, not a replacement — it excels at headless batch tasks where you would not be in the loop anyway.

Want an AI agent without the setup?

Happycapy runs Claude Opus 4.6 in a managed environment — no API keys, no config, just a browser. Starting at $17/mo.

Try Happycapy Free →

Limitations to Know

Goose has real limitations at launch:

No sandboxing by default — it runs commands with your user's permissions. Run it in a Docker container for untrusted repos.
Context window ceiling — large monorepos (1M+ tokens) can exceed model limits. Use a 1M-context model like Gemini 3 Pro for those cases.
No GUI — terminal-only for now. A VS Code extension is on the roadmap but not yet available.
Early maturity — the plugin API may change between versions. Pin your version in production environments.

Why This Matters

Goose is a signal that the era of "AI as suggestion" is ending. The next dominant paradigm is "AI as executor" — agents that do the work, not just recommend it.

Block's decision to open-source it is strategically interesting: they are not trying to monetize the agent directly. They benefit from a stronger developer community, faster iteration on the tool they use internally, and brand positioning as an AI-forward engineering company.

For individual developers, Goose is the most credible free alternative to Claude Code and OpenHands today. The model-agnostic design means your investment in learning Goose does not depend on any single AI provider — a meaningful advantage as the model landscape continues to shift rapidly.