Wednesday, May 20, 2026
S&P 500 · NDX · NVDA · Gold · BTC
  Tool · Agent Platforms All tools →
Block

Goose.

Block's open-source agent framework — MIT licensed, MCP-native.

Category
Agent Platforms
Vendor
Block
Domain
block.github.io
Pricing
Free open source · MIT licence · Self-hosted only
Region
San Francisco, US
Models
Claude, GPT-5, Gemini, Llama · Any LiteLLM-compatible provider

What it is.

Goose is Block's open-source AI agent. Shipped in early 2025 under MIT licence with native Model Context Protocol support, it became the reference MCP client outside Anthropic. Runs locally as a desktop app or CLI, routes to any major model, and uses MCP servers as the standard tool primitive. Block uses it internally across Square, Cash App, and Tidal engineering teams.

Where it fits.

Engineering teams that want a production-grade open-source agent without committing to the LangChain ecosystem. The MCP-native architecture means tool integrations from Anthropic, Block, and the open-source community work out of the box. Strong for developer-tooling and internal automation use cases.

Use cases
  • Developer workflow automation
  • Code generation and refactoring
  • MCP server orchestration
  • Internal tools and DevOps automation
Strengths
  • MIT licensed — fully permissive for commercial use
  • MCP-native — broadest tool ecosystem outside Claude Desktop
  • Backed by Block engineering with documented internal use
Considerations
  • Desktop-first — multi-user hosting story still developing
  • Smaller community than LangChain or AutoGen

Frequently asked.

Why did Block ship Goose?

Block needed an agent across Square, Cash App, and Tidal engineering. The team open-sourced what they were building internally rather than buying a third-party framework.

What is MCP?

Model Context Protocol — Anthropic's open standard for connecting agents to tools and data. Goose was among the first non-Anthropic clients to adopt it.

Is Goose production-ready?

Yes inside Block. External production deployments exist but the framework is younger than LangChain or AutoGen — adopters should expect rougher edges.