Goose is Block's (Square/Cash App) open-source AI agent — a fully autonomous developer agent that goes beyond code suggestions to install dependencies, execute code, run tests, and interact with external services. Notable as a founding member of the Linux Foundation's Agentic AI Foundation alongside MCP and OpenAI's AGENTS.md.
Why It's in Trial
Goose brings enterprise credibility and strong open-source governance to the agentic coding space. With 29,400+ GitHub stars and 368+ contributors since its January 2025 launch, it has earned serious traction. Key differentiators:
- MCP-native: Goose was designed around Model Context Protocol from the start, meaning it connects to any MCP-compatible data source or service — Figma, GitHub, databases, internal tools — without custom integration code. This is its strongest architectural advantage over tools built before MCP became standard.
- Multi-model support: Works with any LLM (Claude, GPT, Gemini, local via Ollama) and can use multiple models in a single session, optimising cost vs capability per task.
- Dual interface: Available as both a CLI and a desktop application — lower friction for developers new to agentic tools.
- Extensible: Extensions (plugins) let teams add organisation-specific tools and integrations without forking the framework.
- Apache 2.0: Permissive licence, appropriate for enterprise adoption.
The MCP Angle for Engineering Managers
Model Context Protocol (MCP) is an open standard that allows AI agents to connect to data systems through a consistent interface — like an "API for AI tools". Goose's first-class MCP support means your organisation's MCP investments (internal tools built as MCP servers) work automatically with Goose. This matters as MCP becomes standard infrastructure.
August 2026: Control-Plane Expansion
Goose 1.46.0 is a broad operating-model release: PreToolUse hooks can deny actions, Summon delegates asynchronous work with explicit context and working directories, and open plugins can bundle skills and MCP extensions. ACP serving gains TLS, reconnect behavior, pagination, and direct desktop integration; MCP apps carry trusted ownership metadata; a Harbor evaluation runner, richer OpenTelemetry semantics, and per-message cost/cache metrics make the agent easier to measure. The same release adds worktree-aware navigation, bounded code-mode execution and tool output, update-provenance checks, and numerous approval, credential, and path-containment fixes. (v1.46.0)
The release appears to aggregate a large backlog rather than one narrow capability. Teams should therefore stage adoption: first validate hooks and approval precedence, then plugin/skill provenance, ACP/MCP ownership, and finally delegated background work. Goose remains Trial because the expansion is substantial and inspectable, while built-in execution isolation is still not its primary safety boundary.
Compared to Claude Code and OpenCode
All three are terminal/CLI-first autonomous coding agents. Goose's distinguishing traits are its MCP-first extensibility model and corporate backing from Block — which brings sustained maintenance and enterprise focus.
| Goose | Claude Code | OpenCode | |
|---|---|---|---|
| Licence | Apache 2.0 | Proprietary | MIT |
| LLM support | Any | Claude (primary) | 75+ providers |
| MCP integration | First-class native | Supported | Supported |
| Governance | Linux Foundation AAIF | Anthropic | Community |
Getting Started
# Install
pip install goose-ai
# Or via Homebrew
brew install block/goose/goose
# Run
goose session start
Key Characteristics
| Property | Value |
|---|---|
| Interface | CLI + desktop app |
| License | Apache 2.0 |
| Underlying model | Any (Claude, GPT, Gemini, local) |
| Pricing | Free (BYOK) |
| Sandbox | Host environment (no built-in isolation) |
| MCP support | Native, first-class |
| Provider | Block Inc. / Linux Foundation AAIF |
| GitHub | block/goose |
| Website | goose-docs.ai |