Goose

What it is

Goose was built inside Block as an internal agent and open-sourced from there. In April 2026 Block donated it to the Agentic AI Foundation, a Linux Foundation project, so it now has vendor-neutral governance rather than a single corporate owner. It sits at around 53,000 GitHub stars.

That governance change is the thing to weigh. Every other harness on this board is steered by a company with a model to sell, an editor to sell, or a cloud to sell. Goose is not, and the difference shows up in how the project handles providers: all fifteen-plus of them are first-class, and none of them is the house choice.

How the loop works

The core is Rust, and there are three surfaces onto the same agent — a desktop application, a command-line interface and an API. The desktop app is unusual company for a developer agent; it makes Goose usable for people who are automating work that is not strictly coding, which is the reason we list it as a general-purpose agent rather than a coding one.

Tools come from extensions, and extensions are MCP servers. Goose was designed alongside Anthropic's Model Context Protocol launch rather than retrofitted to it, so MCP is the native extension mechanism rather than a compatibility layer bolted on later. In practice that means anything in the MCP ecosystem is a candidate tool without adapter work.

Model choice is genuinely open: fifteen or more providers, including local runs through Ollama. If you want an agent that never sends a token to a hosted API, that is a supported configuration, not a hack.

Isolation and permissions

Goose runs on your machine with your permissions. There is no sandbox of its own, so the isolation story is whatever you build around it — a container, a VM, a dedicated user account. Extensions widen the blast radius in the way you would expect: an MCP server can do whatever it was written to do, and adding one is a trust decision.

Who it is for

Organisations that need to justify a tool choice on governance grounds as well as technical ones. A Linux Foundation project with an Apache-2.0 licence answers procurement questions that a single-vendor tool does not, and the provider-agnostic design means a model contract ending is a config change rather than a migration.

It also suits anyone running local models seriously, since Ollama support is treated as a real path rather than a demo.

Limits

Neutral governance costs pace. A foundation project with multiple stakeholders does not ship the way a single company shipping its own agent does, and through 2026 Goose has visibly trailed the fastest-moving harnesses on new capabilities. If you want whatever landed last week, this is not where it lands first. The trade is deliberate: you get stability and no vendor lock-in, and you give up being on the edge.

Alternatives on this board

• OpenCode — also provider-agnostic, narrower in scope, terminal only.
• OpenClaw — MIT, model-agnostic, closer to Claude Code's ergonomics.
• Claude Code — the single-vendor comparison, and the faster-moving one.

Sources

• block/goose on GitHub
• Goose documentation
• Agentic AI Foundation (Linux Foundation) — donation announcement linked from the project repository; we have not verified a separate foundation URL here