Goose

Goose is an open-source AI agent that runs on your own machine as a desktop app, CLI, or programmable API. It was originally built at Block and has since moved to the Agentic AI Foundation at the Linux Foundation for vendor-neutral governance, and it has grown into one of the larger community-driven agent projects (45k+ GitHub stars, 500+ contributors). Where hosted assistants lock you into a single model and toolset, Goose is deliberately provider-agnostic: it talks to 15+ LLM backends including Anthropic, OpenAI, Google, Azure OpenAI, AWS Bedrock, OpenRouter, and local models through Ollama, and it hosts its capabilities through the Model Context Protocol so any MCP server (databases, browsers, filesystems, cloud APIs, internal tools) can be attached as an extension. There are 70+ ready-made extensions and you can write your own. Typical workflows include multi-step coding tasks (planning, edits across a repo, running tests, opening PRs), research assignments that stitch together web fetch, scraping, and note-taking, data-analysis loops against local files or warehouses, and internal automations like triage or report generation. The 'Recipes' system lets you capture a working session as a portable YAML file so a workflow can be re-run, shared, or scheduled. Subagents let a parent agent spawn parallel workers for fan-out tasks, and interactive UI blocks let a recipe surface a small form or chart instead of a wall of text. Security features include prompt-injection detection on tool output and an adversary-review mode that critiques the agent's own plan before it executes destructive actions. It fits developers and small technical teams who want a Cursor/Claude-Code-style agent they fully control, can point at private data without shipping it to a SaaS, and can swap between frontier and local models depending on cost and sensitivity.