Block's goose: A Regulated Fintech Bets on an Open, Model-Agnostic Agent¶
Block (formerly Square) is a regulated financial-services company — not the profile most people picture betting early on an open-source AI agent framework. In January 2025, Block's Open Source Program Office did exactly that, releasing "codename goose" publicly rather than building an internal, closed tool.
What they built¶
goose is a real, open-source (Apache-2.0) AI agent framework, built in Rust, designed to be model-agnostic rather than tied to one provider, and connecting to external capabilities through MCP — Block's own choice to build on an open, standardized tool-connection protocol rather than a proprietary integration layer. As of this writing, the project has drawn 54.6k GitHub stars and has since moved to community governance under the Agentic AI Foundation at the Linux Foundation — the same governance body Skills, A2A and Other Protocols covers for AGENTS.md, a pattern worth noticing: several agent-ecosystem projects that started as one company's internal tool have moved to neutral, cross-vendor governance once adoption grew past that one company.
Why, in Block's own words¶
Block's own open-source lead, Manik Surtani, framed the decision as being about more than developer goodwill: open source is the "best way to build software," and — the more specific, company-relevant point — "even more important for a financial services company to be building on open principles and open standards." For a regulated fintech, the more usual instinct is to keep infrastructure proprietary and tightly controlled; Block's own stated reasoning runs the other way, treating openness and auditability as the more defensible choice given the regulatory stakes, not a luxury a less-regulated company could better afford.
A real, necessary caveat: several secondary sources circulate claims that goose "replaced 40% of Block's engineering team" or "scaled to 60% of the company" — these numbers could not be verified against any primary Block source and read as unsourced content-farm amplification. They're worth flagging explicitly as not usable rather than silently repeating them, the same standard this site holds every other claim to.
The lesson
A model-agnostic, MCP-native architecture is a real, deliberate hedge against vendor lock-in specifically — goose's own design means Block isn't betting its internal tooling on any single model provider's roadmap, which matters more, not less, for a company operating under regulatory scrutiny where switching infrastructure later isn't a quick decision. The broader, more general lesson: when a framework choice claims real numbers, check whether they trace to the company itself or to secondary "AI news" amplification — this case study's own real, verified facts (the star count, the Linux Foundation governance move, Surtani's own quote) are worth far more than the unverifiable adoption-percentage claims that get repeated alongside them.