Goose is an open-source AI agent for desktop and terminal, developed by Block and handed over to the Linux Foundation’s Agentic AI Foundation in December 2025. As an AI harness, Goose connects a freely chosen language model to files, shell and tools through MCP extensions. As of 22 November 2026.
What is Goose?
Goose is a general-purpose agent for programming, research, writing, automation and data analysis, available as a desktop app and as a CLI for macOS, Linux and Windows. Block, the company behind Square and Cash App, released version 1.0 on 28 January 2025 according to GitHub. Per the Linux Foundation press release of 9 December 2025, Goose is one of the three founding projects of the Agentic AI Foundation (AAIF), alongside MCP and AGENTS.md. The repository now lives at aaif-goose/goose; per the GitHub API, the old path block/goose redirects there.
The project is very active: per the GitHub API it has 56,867 stars, the current version per the releases is 1.60.2 from 20 November 2026, and eight versions shipped between 9 and 20 November 2026 alone (as of 22 November 2026). You install the CLI with the install script from the README and the desktop app from the download page in the documentation.
Goose as an AI harness
Goose covers all seven building blocks of a harness itself. Tools come from extensions that, per the documentation, are built on the Model Context Protocol; any MCP server can be added as an extension. The agent loop runs interactively in the app or with goose session, headless with goose run for scripts and CI, and as an ACP server (goose acp), which the Zed editor, for example, uses to launch Goose as its agent. Goose reads project rules from AGENTS.md first, then from .goosehints.
Per the documentation on permission modes, three modes control autonomy: Autonomous (default, no prompts), Manual Approval (confirmation before tools) and Smart Approval (risk-based approval). Subagents handle subtasks in parallel but are only active in Autonomous mode. Recipes bundle extensions, prompts and settings into reusable workflows; hooks, skills and plugins follow open specifications. If you want to build your own agents, the Rust library oi-kit is available, which per the README is still in alpha.
Models and model access
Goose is model-agnostic and, per its README, connects to more than 85 providers, including Anthropic for Claude, OpenAI for GPT, Google for Gemini, Amazon Bedrock, Azure OpenAI, OpenRouter and Databricks. Local models run via Ollama, LM Studio or directly inside the Goose process without a separate model server. Through the Agent Client Protocol, Goose can also use Claude Code or Codex with their subscriptions as providers.
Custom endpoints are explicitly supported per the provider documentation: the OpenAI provider connects to any OpenAI-compatible service via OPENAI_HOST and OPENAI_BASE_PATH, and OPENAI_CUSTOM_HEADERS sets custom headers. For multiple endpoints you create custom providers in OpenAI, Anthropic or Ollama format, which can be shared across the team as JSON files. This lets you run Goose in proxy mode through the AI Gateway, which checks permissions and budgets and assigns every token to a cost center.
Enterprise use: data protection and governance
Goose runs locally and sends prompts and code only to the chosen model provider; per the documentation, that provider’s retention and privacy policies then apply. Goose stores sessions in a local SQLite database. Per the telemetry documentation, Goose sends anonymous usage data only after consent on first launch; conversations and code are not included.
For organizational control, Goose offers an extension allowlist that is loaded from a central URL via GOOSE_ALLOWLIST, as well as custom distributions with preconfigured providers and extensions. Check the default mode: Autonomous changes and deletes files without asking, and the documentation does not describe a built-in sandbox. Set Manual or Smart Approval as the default, enable prompt injection detection and Adversary Mode where needed, and isolate risky tasks in Docker. Regardless of that, source code and prompts belong in your record of processing activities.
Running Goose with herdr and Orca
Goose is not currently supported by herdr: the herdr agent list in version 0.9.5 does not name Goose. Orca lists Goose in its own agent overview with automatic setup, without deeper integration such as status reporting or hooks. There you can start several Goose sessions in parallel, each in its own Git worktree.
How does Goose differ from Claude Code and Codex?
Goose is vendor-neutral, whereas Claude Code and Codex come from the model vendors Anthropic and OpenAI and are tailored to their models and subscriptions. Goose is governed by a Linux Foundation project foundation, offers a desktop app alongside the CLI and also targets tasks beyond code. OpenCode is similarly open in model choice but focuses on the terminal. One important difference in daily use: Goose starts in Autonomous mode without a sandbox, Claude Code asks before file changes and shell commands in its default mode, and Codex limits actions through a sandbox with an approval policy.
Where it fits: developer harness and enterprise harness
Goose is a harness for developers and technically skilled teams. For business users who work without a terminal, CompanyGPT is the matching harness, with MCP tools, enterprise knowledge and roles from Entra ID or Keycloak. The AI Gateway provides the shared cost and governance layer; per the product page, its companion measures agent time for Claude Code, Codex, OpenCode, Pi, GitHub Copilot CLI and Gemini CLI, so you capture Goose through proxy mode. Our overview of AI harnesses covers further coding agents and orchestrators.
