What it is
goose is an open-source AI agent that runs on your computer as a desktop app for macOS, Linux and Windows, a CLI, or an embeddable API. It can write and run code, edit files and automate other tasks, and it is built around MCP extensions, with more than 70 listed.
Block created it and donated it to the Agentic AI Foundation (part of the Linux Foundation) in December 2025; the repository moved from block/goose to aaif-goose/goose in April 2026. It works with more than 15 providers, including Anthropic, OpenAI, Google, Azure, Amazon Bedrock, OpenRouter and local models through Ollama, and can use an existing Claude, ChatGPT or Gemini subscription through ACP.
Our review
goose is the open agent to choose when the work spans more than a code editor, and the one whose defaults most need changing before you trust it. It has about 55,600 GitHub stars and shipped v1.55.0 on 16 October 2026. Its move to the Agentic AI Foundation, alongside MCP and AGENTS.md, gives it governance no single vendor controls, which matters to companies wary of building on a start-up's roadmap.
Its breadth comes from MCP. Recipes package a prompt, extensions and settings into something a colleague can rerun, scheduled tasks run them unattended, and custom distributions let a company ship its own build with preset providers and branding. The desktop app makes this usable for analysts who never open a terminal.
The defaults are permissive. The docs say Autonomous mode, in which goose edits and deletes files and calls extensions without asking, is applied by default; Manual, Smart and Chat-only modes exist but must be chosen. In headless runs the Auto mode approves any tool confirmation on its own. A high-severity advisory in July 2026 (CVE-2026-72718) showed the risk: goose review could run arbitrary commands through a repository's git core.fsmonitor setting, fixed in 1.44.0.
As a pure coding agent it trails OpenCode and Codex on coding-specific features, and quality swings with the model and extensions you configure; the known-issues page lists loops in long sessions and hangs on commands that never exit, such as npm run dev. Telemetry, by contrast, is opt-in and limited to counts, versions and error types.
There is no hosted goose, so the total cost is your model bill, which a long autonomous session with a dozen extensions loaded can run up faster than a chat assistant on a fixed subscription.
Where it earns its place
An operations analyst at a 50-person agency
Pull weekly figures from Google Drive, a CRM and Slack into one report
A recipe bundles the prompt and the three MCP extensions, and a scheduled task runs it every Monday in the desktop app. goose is free; the cost is model tokens, and Manual or Smart approval should be set before it touches shared files.
A platform team at a 2,000-person company
Give staff a pre-approved internal AI agent
Build a custom distribution with the company model gateway, an allowlist of extensions and the default mode set to Smart Approval, then roll it out like any desktop app. The licence (Apache 2.0) permits rebranding; support is community only.
A backend developer on a privacy-sensitive codebase
Run a coding agent with no code leaving the laptop
Point goose at a tool-calling model in Ollama such as qwen2.5 (the docs note DeepSeek models cannot call tools) and leave telemetry off. Expect weaker results than a frontier API model on multi-file changes.
Before you commit
- Setup
- Download the desktop app or install the CLI with the curl script from the README, then run goose configure to pick a provider; a first session takes minutes once an API key or Ollama model is ready.
- Integrations
- More than 15 model providers, Claude, ChatGPT and Gemini subscriptions via ACP, more than 70 MCP extensions, and an API for embedding goose in other tools.
- Data and privacy
- Runs locally; prompts go only to the provider you configure. Anonymous usage data is opt-in and excludes conversations and code. Autonomous mode is on by default. One published advisory, CVE-2026-72718 (high), fixed in 1.44.0.
- Cost at scale
- No paid edition exists; cost is entirely model usage, or local GPU hardware with Ollama. Long autonomous sessions and large extension sets raise tokens per task.
- Limits
- About 514 open issues and pull requests. Known issues include loops in long sessions, hangs on non-exiting commands, context overflow, and keyring problems in containers and headless systems.
What's good
- Neutral Linux Foundation governance
- More than 70 MCP extensions, plus recipes and schedules
- Desktop app, CLI and API
- Opt-in telemetry; works with local models
What's not
- Autonomous mode is the default
- Fewer coding-specific features than OpenCode or Codex
- Quality depends on the model and extensions chosen
- No managed or hosted option
Best for
Developers and power users who want an open, extensible agent for coding and general automation.
Alternatives
goose: common questions
- Is goose worth it?
- A vendor-neutral agent for automating work across many systems, with coding as one job among several and autonomy switched on from the first prompt. We rate it 4.1 out of 5.
- How much does goose cost?
- Free and open source (Apache 2.0). You pay your model provider or run local models.
- What are the downsides of goose?
- Autonomous mode is the default; fewer coding-specific features than OpenCode or Codex; quality depends on the model and extensions chosen; no managed or hosted option.
- Who is goose best for?
- Developers and power users who want an open, extensible agent for coding and general automation.
- How much does goose cost for a team or in production?
- No paid edition exists; cost is entirely model usage, or local GPU hardware with Ollama. Long autonomous sessions and large extension sets raise tokens per task.
- Is goose safe for company data?
- Runs locally; prompts go only to the provider you configure. Anonymous usage data is opt-in and excludes conversations and code. Autonomous mode is on by default. One published advisory, CVE-2026-72718 (high), fixed in 1.44.0.
- What are the best goose alternatives?
- The closest alternatives we have reviewed are OpenCode, OpenHands and OpenAI Codex CLI.
Answers checked Oct 23, 2026.
Make goose? Show this rating on your site
