Reviewed by Harsh Desai · Last reviewed:
Goose
Open-source AI coding agent that runs locally with any LLM
Best for
Goose is the open-source AI coding agent built by Block Inc., the company behind Square and Cash App. Released under Apache-2.0 with 44,000+ GitHub stars and 4,500+ forks, it runs entirely on your machine and connects to any large language model through an OpenAI-compatible endpoint. The agent can read your codebase, install dependencies, edit files across folders, run the test suite, and open pull requests without leaving the terminal.
What Goose does:
- •Autonomous task execution give Goose a goal and it plans the steps, edits multiple files, runs tests, fixes failures, and opens a pull request without manual hand-holding between steps.
- •Bring-your-own-LLM connect to Claude, GPT, Gemini, or any OpenAI-compatible endpoint, plus local Ollama, LM Studio, and llama.cpp servers for fully offline workflows.
- •3,000+ MCP integrations ships with Model Context Protocol support, with extensions for GitHub, Slack, Jira, Docker, browser automation, and IDE control. Each extension is malware-scanned before activation.
- •Desktop GUI plus CLI a native desktop app on macOS, Linux, and Windows for non-terminal users, alongside a full CLI for automation and scripting.
- •Self-hosted by design the entire agent runs locally, so source code, prompts, and tool calls never leave your machine when paired with a local model.
- •Recipes for repeatable workflows save a multi-step task as a recipe (lint then test then commit then PR) and run it across projects with one command.
- •Built and used inside Block Block engineers use Goose daily for production work on Square and Cash App, which keeps the roadmap focused on real developer pain rather than demo features.
- •Extensible by writing your own MCP server add a private tool (internal API, custom build script, proprietary database) by writing a small MCP server in any language.
- •API-first architecture Goose exposes its agent loop programmatically, so you can embed it in CI runners, code review bots, or custom automation pipelines.
- •Chat history and session resume pause a long-running task, close the laptop, and resume the same context days later without losing the agent's mental model.
Pricing:
- •Goose core $0: free forever under Apache-2.0. No subscriptions, seat limits, or feature paywalls.
- •LLM provider costs $0 to ~$50/month: pay your chosen LLM provider directly (Claude, GPT, Gemini) or run Ollama locally for $0.
- •Self-hosting $0 to ~$10/month: a small VPS, a workstation, or a Mac mini handles individual or small-team usage.
Limitations:
- •Setup is more technical than Cursor installing Goose, configuring an LLM provider, and learning the recipe system takes longer than opening a polished IDE.
- •Output quality depends on the LLM you pick pair Goose with a weak local model and results will lag Claude Code, even though the agent loop is the same.
- •Discovery across 3,000+ MCP extensions takes time the catalogue is large, the official curation is light, and finding the right extension for a niche task can be a research task in itself.
Our Verdict
Goose scores 8.7/10 because it gives developers a fully open-source, self-hostable AI coding agent that handles complete workflows without locking you into a single model or a $200-per-month subscription. With 44,000+ GitHub stars, 4,500+ forks, and daily production use inside Block (Square, Cash App), the project has the engineering depth that drive-by AI tools lack, and the Apache-2.0 license means you actually own what you deploy.
For the Vibe Builder, Goose is the answer when ChatGPT-style copy-paste stops scaling and you need an agent that can edit ten files, run the tests, and open a pull request without your help. Pair it with Claude or GPT for serious work, or with a local Ollama model for free experimentation, and you get production-grade autonomy on a $5 VPS or a laptop. The desktop app makes the entry path softer than a pure CLI, and recipes turn one-off prompts into reusable team workflows.
For the Developer, the value is full ownership. The agent loop, the prompts, the tool calls, and the MCP extensions are all in a readable Rust codebase you can fork, audit, or extend. Writing a private MCP server takes an afternoon, embedding Goose in CI takes a config file, and there is no vendor that can change pricing or rate-limit your team out of nowhere. The 3,000+ MCP catalogue covers GitHub, Slack, Jira, Docker, and most internal-tool patterns out of the box.
Skip it if you want a polished, zero-setup IDE and would rather not configure an LLM provider yourself. In that case consider Cursor for the editor experience, or try Claude Code if you prefer a managed agent with Anthropic models built in by default.
Related Tools
View allCompare Goose With
Also Useful For
Frequently Asked Questions
How much does Goose actually cost to run in 2026?
Goose itself is free forever under Apache-2.0. The real cost is the LLM you connect to. Cloud models like Claude or GPT typically run $10-50 per month for individual use. Ollama or LM Studio with a local model is $0. Many small teams run Goose on a $5/month VPS plus a metered Claude key.
Goose vs Claude Code: which agent should I pick?
Goose is open-source, model-agnostic, and free, but you configure the LLM yourself. Claude Code is a managed Anthropic product with Claude models built in and a polished CLI. Choose Goose when you want full ownership, model choice, and zero subscription. Choose Claude Code when you want a single bill, no setup, and Anthropic-tuned defaults.
Which large language models can Goose connect to?
Anything with an OpenAI-compatible API. Block officially documents Claude (Anthropic), GPT (OpenAI), and Gemini (Google) for cloud, plus Ollama, LM Studio, and llama.cpp for local. Custom endpoints like LiteLLM, vLLM, Together AI, and Fireworks work without code changes once configured.
Is Goose safe to use on a real production codebase?
Goose runs locally and Block Inc. engineers use it daily on Square and Cash App, which is a strong production signal. Every MCP extension is malware-scanned before activation. Standard practice is to run Goose against a feature branch, not main, and review the pull request it opens before merging.
How do I install Goose on Windows, Linux, and macOS?
macOS supports a native desktop app, Homebrew, and a curl install script. Linux supports DEB packages for Ubuntu and Debian, RPM for RHEL and Fedora, and Flatpak. Windows has a dedicated installer plus a CLI build. All paths take a few minutes and install both the agent and the MCP runtime.
What is Goose?
Goose is Open-source AI coding agent that runs locally with any LLM.
Is Goose free?
Yes, Goose offers a free version.
Who should use Goose?
Goose is built for vibe builders who want AI to handle the technical work and developers looking to accelerate their workflow. Common use cases include autonomous-coding, open-source-agent, local-llm-coding, mcp-automation, team-coding-workflows.
What are the best alternatives to Goose?
Popular alternatives to Goose include Cursor, Claude Code. Compare features and pricing in our Coding directory to compare options.
Affiliate link: we may earn a commission. How this works.
Goose
Free tier available