Goose Review: Block's Open-Source AI Agent, Now Under the Linux Foundation

Run an agentic coding session from the CLI

Install the CLI, run goose session from inside a project directory, and describe a task in plain language (for example, "add input validation to the signup form and write a test for it"). Goose reads the relevant files, makes the edits, runs the test suite, and reports back — asking for confirmation on higher-risk actions depending on your configured permissions.

Connect goose to a fully local model via Ollama

With Ollama installed and a model pulled locally, configure goose to use Ollama as its provider in the desktop app's or CLI's provider settings. Goose then runs entirely against your local model, with no API key and no data leaving your machine — at the cost of whatever that local model's capability ceiling is compared to a frontier cloud model.

Add an MCP extension for a specific tool

Browse goose-docs.ai/extensions for an MCP server matching a tool you use (GitHub, a database, a browser), install it through the desktop app's Extensions panel or the CLI, and goose gains the ability to call that tool directly as part of a task — for example, opening a pull request once a code change passes tests.

Use goose if

• You want a free, open-source AI agent governed by a neutral foundation (the AAIF at the Linux Foundation) rather than a single vendor
• You want to choose your own LLM provider — including a fully local, no-cost Ollama model — instead of being locked to one built-in model
• You want a desktop app, CLI, and API on the same underlying agent engine, so a workflow is not locked to one interface
• You want to extend the agent with Model Context Protocol (MCP) tools and reusable YAML "recipes" for repeatable team workflows
• You are comfortable configuring permissions and reviewing agent actions rather than relying on a mandatory sandbox

Wait, or look elsewhere, if

• You want every agent run isolated inside a sandboxed container by default with no configuration — OpenHands sandboxes by default; goose's default is closer to your own machine, so review its permission settings before running untrusted tasks
• You want a single built-in model with nothing to configure — goose's bring-your-own-key model means you must pick and connect a provider before you can use it at all
• You need official vendor support with an SLA rather than community and foundation support channels — goose's support is GitHub Issues and community channels, typical of an open-source foundation project rather than a commercial vendor
• You want an IDE-embedded coding assistant rather than a standalone agent — an editor extension like Cline may fit your workflow better

Mistake 1: Assuming goose includes a built-in model or API key

Goose is bring-your-own-LLM: it has no built-in model of its own. You must connect it to a provider (Anthropic, OpenAI, Google, Ollama, and dozens more) with your own API key, or a local Ollama model, before it can do anything.

Mistake 2: Linking to the old block/goose GitHub URL

Goose's repository moved to github.com/aaif-goose/goose on April 7, 2026, as part of the donation to the Agentic AI Foundation. The old block/goose URL redirects, but bookmarks, scripts, and documentation should be updated to the new canonical location.

Mistake 3: Treating Block as goose's current sole governing entity

Block built and launched goose, but as of April 7, 2026, project governance sits with the Agentic AI Foundation (AAIF) at the Linux Foundation — a neutral body Block co-founded alongside Anthropic (MCP) and OpenAI (AGENTS.md), not a Block-only project.

Mistake 4: Assuming a mandatory sandbox

Unlike some competitors (OpenHands, for example) that isolate every agent run inside a Docker container by default, goose runs closer to your own machine unless you configure otherwise. Review its permission and confirmation settings before pointing it at a task with real file or shell access.

Mistake 5: Assuming stale star counts or license details

This review verified 54,156 GitHub stars and an Apache 2.0 license directly via the GitHub API on September 12, 2026 — re-check that endpoint yourself for a current figure rather than relying on any cached number, including this one.