About 60% of Block's 12,000 employees use Goose weekly. That is an unusually candid internal adoption number for an open-source project - and it puts some context around why Block donated the thing to the Linux Foundation rather than spinning it into a product.
In December 2025, Block contributed Goose as a founding project of the Linux Foundation's Agentic AI Foundation (AAIF), alongside Anthropic's MCP and OpenAI's AGENTS.md. The project's repository and governance were transferred from Block to the AAIF in April 2026. That transfer is the signal worth understanding. It means Goose is no longer a company project with an open-source veneer - it is now a foundation-governed project with the same structural independence as Linux or Kubernetes. Whether you trust that framing is a separate question, but the governance is real.
What Goose actually does, and what makes it different
Goose can read and modify files, execute shell commands, build and run software, perform tests, and interact with external services through an agent harness. It is available as a native desktop application for macOS, Linux, and Windows, as well as through a CLI and an API. That covers the same surface as Claude Code or Codex CLI, so the meaningful differentiators are elsewhere.
The first is model agnosticism. Goose is model-independent and supports multiple AI providers, including Anthropic, OpenAI, Google, Microsoft Azure, Amazon Bedrock, OpenRouter, and locally with Ollama.
A single configuration change swaps the underlying model from Claude to GPT-4o to Gemini to a locally-hosted Llama variant, with no code changes required. For a team that wants to route sensitive code through a local model and let a frontier API handle complex refactors, this is not a minor feature - it is the architecture.
The second is the license. The Apache 2.0 license allows commercial use, modification, and embedding without source disclosure obligations, making Goose viable for companies building their own internal tooling on top of the agent framework.
The third is the extension surface. Goose supports 25+ model providers including Anthropic, OpenAI, Google Gemini, Bedrock, and Vertex, plus local inference via Ollama and Docker Model Runner. It includes 70+ documented extensions and MCP Apps that can render interactive UIs inside Goose Desktop.
The BYOK cost math - and where it breaks down
Goose has no subscription fee. It is fully free and open-source under Apache 2.0 with no paid tier. Goose follows a BYOK model: users pay only for the LLM API usage they choose.
At team scale, that math is compelling. A team of 10 developers might spend $200/month on API tokens through Goose versus $2,000+ on Claude Code subscriptions. That is roughly a 90% reduction for moderate-use teams.
But there is a real cost hiding in the comparison. BYOK API costs require active management: users must monitor LLM API spending across providers, configure billing limits, and handle rate limit errors themselves - more operational overhead than fixed-subscription tools that bundle API costs into a flat monthly fee. That overhead is not hypothetical. A team that plugs in Claude Sonnet via API and runs several long agentic sessions a day can exceed a flat Claude Code subscription before they notice.
The honest read: Goose wins on team-scale deployments and for shops that want to mix providers. Individual heavy users who want the best model performance with the least configuration should run the per-token numbers before assuming they save money.
Where the governance move matters for teams
The Linux Foundation transfer is worth taking seriously, not just as a press event. The AAIF provides a neutral, open foundation to ensure agentic AI capabilities evolve transparently, collaboratively, and in ways that advance the adoption of leading open source AI projects.
What that means practically: Goose now has the same kind of multi-vendor governance backstop that kept projects like Kubernetes from being quietly steered by a single company's roadmap. Platinum members include Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft, and OpenAI. When eight large vendors are co-governing the standard, no single one can deprecate an integration you depend on without community review.
For regulated or privacy-sensitive teams, the local inference path is also meaningfully documented. Configure Goose with an Ollama provider and a locally-hosted model to operate completely offline - no API keys required, no external network calls. The Ollama server runs locally; Goose connects to it via localhost. This is suitable for air-gapped environments and privacy-sensitive codebases.
Beagle in action#dev-tools, 11:02amThe security disclosure you should read
In January 2026, Block's Detection and Response Team published a detailed account of how they red-teamed Goose and successfully ran a simulated infostealer against a developer laptop. Block's red team successfully compromised a developer's laptop, then fixed the vulnerabilities.
Block's DART team caught the simulated attack and contained it. Goose runs code locally, which limits the blast radius of a compromise to your machine.
The honest point here: that an agent that can read files, execute commands, and call external services is a prompt-injection target is not surprising. What is unusual is publishing the attack chain before a reporter finds it. Block's red team exercise found and fixed prompt injection vulnerabilities in early 2026.
For teams evaluating Goose against proprietary tools, the security comparison is not "Goose is safer" or "Goose is riskier." Goose runs code locally, which limits the blast radius of a compromise to your machine. Claude Code sends code to Anthropic's servers, which introduces a different threat model. Neither approach is inherently more secure. Evaluate against your actual threat model: local execution risk versus cloud-side data exposure.
Goose open-source coding agent: common questions
What is Goose and who maintains it?
Goose is a free, Apache 2.0 open-source coding agent originally built by Block. In April 2026, governance transferred to the Agentic AI Foundation at the Linux Foundation. Block continues as an active maintainer, but the project is now community-governed. It runs on macOS, Linux, and Windows as a desktop app, CLI, and embeddable API.
How does Goose compare to Claude Code on cost?
Goose itself is free; you pay only for LLM API tokens. A 10-person team might spend ~$200/month on API usage versus $2,000+ on Claude Code subscriptions. Individual heavy users can close that gap fast, though - API costs at high volume can exceed a flat subscription. Always run the per-token numbers for your usage pattern before switching.
Can you run Goose with a local model and no internet?
Yes. Configure Goose with Ollama and a locally-hosted model - Qwen3 Coder, DeepSeek Coder, and others all work. No API keys, no external network calls. Goose connects to Ollama via localhost, making it usable in air-gapped or compliance-restricted environments where no code can leave the organization's infrastructure.
Is Goose safe to use on a production codebase?
It carries the same risk profile as any agent with shell access. Block's own security team red-teamed it in early 2026, found prompt injection vulnerabilities, and published the findings - which is more transparency than most vendors offer. Running with a local model reduces the data exposure surface. Apply the same caution you would to any tool that can read files and execute commands.
What is the difference between Goose and OpenHands or Aider?
All three are open-source coding agents with shell and file access. Goose's distinguishing features are its Apache 2.0 license (commercial-friendly with no source disclosure requirements), its native desktop app, its 70+ MCP extensions, and its Linux Foundation governance. OpenHands is stronger on SWE-bench evaluations for pure coding tasks. Aider is leaner and better suited to git-commit-level workflows with tight token budgets.