Block — the company behind Square, Cash App, and Afterpay — didn’t build Goose to sell you a subscription. It built Goose because it needed a code agent for its 12,000 employees that wouldn’t lock them into a single AI vendor. The result today is open source, has 37K stars on GitHub, and is probably the most flexible code agent tool available right now.
The pitch is simple: a local, autonomous agent that can build projects from scratch, write and execute code, debug errors, orchestrate workflows, and call external APIs — with whatever LLM you choose, including local models.
What makes Goose different
Most AI coding tools are vertically integrated. Cursor runs its own models. Claude Code runs on Anthropic’s API. GitHub Copilot runs on OpenAI’s. They pick the model, and the price reflects it.
Goose flips this completely. You connect the LLM you prefer — Claude, GPT-4o, Gemini, a local model with Ollama — and Goose handles the agent orchestration. You can even set up multi-model setups: use Claude Sonnet for architecture decisions, a cheaper Gemini Flash for routine refactoring, and a local llama.cpp model for sensitive code you don’t want leaving your machine.
The privacy angle is real, not marketing: with local models, nothing leaves your machine. No API calls, no cloud processing. And since everything’s under the Apache 2.0 license, every line of code is auditable.
Recipes: the concept that turns individual knowledge into team knowledge
Every other agent relies on prompts — knowledge that lives in your head, or maybe in some note. Goose has Recipes.
A Recipe is a YAML file that packages a complete task: the goal, required extensions, structured inputs, and even sub-recipes for complex workflows. A simple example:
version: 1.0.0
title: Weekly Status Report
description: Generate a status report from Linear, GitHub, and Notion
prompt: |
Generate a weekly status report for {{ engineer_name }}.
Check open issues in Linear, recent commits in GitHub,
and docs updated this week in Notion.
Format it as a Slack-ready summary with links.
extensions:
- linear
- github
- notion
A senior developer writes this recipe once. The whole team runs it with /weekly-status. The prompt is no longer individual knowledge — it’s institutional knowledge, versioned and shareable.
This is something no other major code agent offers at this level of structure.
Getting started in 5 minutes
Installation via CLI:
curl -fsSL https://github.com/block/goose/releases/latest/download/install.sh | bash
Or with Homebrew:
brew install block/tap/goose
Or download the desktop app from block.github.io/goose — it includes a built-in Extensions Manager to discover and enable MCP servers without touching config files.
First run — set up your provider:
goose configure
This guides you through selecting your LLM provider and entering your API key. You can also select Ollama here to run completely offline.
Start a session:
goose session start
From there, describe what you need. Goose takes over — reads your filesystem, executes commands, makes edits, and iterates autonomously until it’s done.
MCP built in from the core
Goose wasn’t retrofitted with MCP support — it was built on MCP from day one (Block co-developed the protocol with Anthropic). Each Goose “extension” is an MCP server.
This means any tool with an MCP server works natively with Goose. To add an extension:
goose configure
# Select: Add Extension
# Enter the MCP server details
You have GitHub, Slack, Google Drive, Jira, Docker, Postgres, and 3,000+ more tools through standardized connections. Recipes have automatic access to globally configured MCP servers — and recipe authors can pin specific servers and tools to ensure reproducibility.
IDE integration via ACP
Goose integrates with code editors through ACP (Agent Client Protocol) — an open standard for connecting AI agents with IDEs, built by Zed Industries and JetBrains.
Currently supported editors:
| Editor | Status |
|---|---|
| Zed |  Supported |
| Neovim | Supported |
| JetBrains (IntelliJ, PyCharm, WebStorm, etc.) | via ACP Registry (since January 2026) |
| Marimo | Supported |
| VS Code |  Not confirmed |
| Cursor | Doesn’t support ACP (as of March 2026) |
| Windsurf | Not confirmed |
For JetBrains users specifically, this is a significant unlock — Goose can be installed with a click from inside IntelliJ, PyCharm, or WebStorm without a separate subscription.
The honest trade-off
Goose’s score on SWE-bench with equivalent backend models is around 45%. Claude Code scores 72.7% on the same benchmark. If pure code quality on hard tasks is your main metric, Claude Code wins.
Where Goose wins: flexibility, cost control, privacy, IDE coverage (especially JetBrains), and the Recipes system for team workflows. And it’s genuinely free — Apache 2.0, no subscription.
The right framing isn’t “Goose vs Claude Code.” It’s: do you need the best possible code quality from a single vendor, or do you need an agent your whole team can standardize on, with different models and environments?
The security story worth knowing
In January 2026, Block published something most companies would’ve buried: its own red team successfully hacked Goose in an exercise called Operation Pale Fire. The attack used a poisoned Recipe with malicious instructions hidden in invisible Unicode characters — tricking both the developer and the agent into downloading an infostealer.
Block’s CISO compared it to autonomous vehicles: it’s not enough to be as safe as humans; they need to be safer. They found it, published it, and fixed it. For an open source project betting that recipes get shared across teams, this kind of security transparency matters.
Who Goose is for
Goose makes more sense for:
- Teams that don’t want lock-in with any LLM vendor
- Developers using JetBrains IDEs who want agentic AI without switching editors
- Anyone with privacy requirements that make cloud code analysis unviable
- Teams that want to encode expert knowledge in shareable, versioned Recipes
- Developers comfortable with configuration who want maximum flexibility
It’s not the right choice if you want a polished point-and-click experience with maximum out-of-the-box accuracy. For that, Claude Code or Cursor remain stronger options.
Goose is at 37K stars and climbing. The open source code agent space is genuinely competitive now — and Goose is one of the most carefully architected options in it.