goose

• Asking goose to scaffold, run, and test a small app or script from the desktop app or terminal
• Packaging a team workflow, such as a code review or release checklist, as a recipe that teammates or CI can run
• Running a recipe every morning on a cron schedule to produce a report
• Splitting a large task across parallel subagents, for example reviewing several modules at once
• Using a local model through Ollama for work that should not leave the machine
• Driving goose from an editor such as Zed through ACP, or embedding it through `goose serve`

goose covers more surfaces than most open-source agents: a desktop app, a CLI, and a server mode. It works with a long list of providers, including subscriptions you already pay for, and recipes turn one-off prompts into workflows you can share and schedule. Foundation governance means no single vendor controls its direction. The cost of that breadth is safety defaults. It runs in autonomous mode out of the box and the security reviewers are opt-in, so teams should choose a permission mode deliberately before pointing it at important repositories.

About goose

goose was built at Block and donated to the Agentic AI Foundation (AAIF) at the Linux Foundation. In April 2026 the repository moved from block/goose to aaif-goose/goose and the docs moved to goose-docs.ai. It is written in Rust and ships as a desktop app for macOS, Linux, and Windows and as a CLI that share the same configuration. You connect a model provider and goose works through tools. The core agent reads and edits files and runs commands, and extensions built on the Model Context Protocol (MCP) add databases, browsers, GitHub, Google Drive, and other systems. The site lists 70+ extensions. goose is not tied to one vendor. It works with API keys for Anthropic, OpenAI, Google Gemini, xAI, Mistral, and others, with cloud platforms such as Amazon Bedrock, Azure OpenAI, and Google Vertex AI, and with local models through Ollama, LM Studio, Docker Model Runner, and Ramalama. It can also run on subscriptions you already pay for: ChatGPT Codex and GitHub Copilot providers, plus ACP providers that drive Claude Code, Codex, Amp, or Pi. Three features go beyond a typical coding agent. Recipes package instructions, extensions, and parameters into YAML files you can share, launch from the desktop app, CLI, deeplinks, or CI, and schedule with cron. Subagents take tasks in sequence or in parallel. And `goose run`, `goose acp`, and `goose serve` let you script it headlessly, connect it to ACP editors such as Zed, or expose it as an HTTP/WebSocket server. The tradeoff is how much it can do by default. Autonomous mode is the default, so goose edits files and runs commands without asking unless you switch to Manual Approval, Smart Approval, or Chat Only. Prompt injection detection and the adversary reviewer are opt-in, and the docs call them a safeguard rather than a guarantee. goose also depends on models with strong tool calling, and releases ship several times a month, so behavior and settings change often.