HARNESSES AND BROKERS / GENERAL-PURPOSE AGENT

goose (AI agent)

goose began at Block and is now part of the Agentic AI Foundation at the Linux Foundation. It is a useful reference for an agent that is built on open standards from the inside: its extensions are MCP servers, and its repeatable workflows are portable YAML recipes.

Technology

What it is

goose is a general-purpose AI agent that runs on your own machine. It ships as a desktop app for macOS, Linux, and Windows, as a CLI, and as an API for embedding. It is licensed Apache 2.0 and written mostly in Rust. The project writes its name in lowercase.

Governance and the Agentic AI Foundation

goose was created at Block. In December 2025 the Linux Foundation formed the Agentic AI Foundation (AAIF) with three founding projects: Anthropic's Model Context Protocol, Block's goose, and OpenAI's AGENTS.md. The repository now lives under the aaif-goose organization with a published governance file.

For an open architecture, neutral governance is part of the openness story. A harness owned by one company can change direction with that company; a foundation project has a process for how decisions get made.

Extensions are MCP servers

goose extends itself through the Model Context Protocol, and the project lists 70+ extensions. The design consequence is that goose has no private plugin format to lock into: an extension built for goose is an MCP server, usable by any other MCP client.

Recipes as portable workflows

Recipes capture a workflow as a YAML file: instructions, the extensions it needs, and parameters. They can be shared across a team and run in CI. A recipe is the goose equivalent of a skill or a saved procedure, kept as a file you own and can version, rather than as settings inside the product.

Model providers

goose works with 15+ providers, including Anthropic, OpenAI, Google, Ollama, OpenRouter, Azure, and Bedrock, and can use existing Claude, ChatGPT, or Gemini subscriptions through the Agent Client Protocol (ACP). Local models through Ollama keep work on your own hardware when that matters.

Beyond coding

goose is positioned for research, writing, automation, and data analysis as well as code. That puts it closer to a personal or operations agent than a pure coding harness, and it means the bounded and grounded questions apply to whatever systems its extensions reach, not only a repository.

Where it fits in an open stack

goose is an execution-layer agent built almost entirely on open interfaces: MCP for tools, many providers for models, YAML recipes for workflows, and foundation governance for the project itself. It is a strong example of the composable pattern this reference describes.

Gotchas worth knowing

  • Every extension is a grant of authority. An MCP server connected to email, files, or a database extends what the agent can do. Review the list like a permission set.
  • Third-party extensions are code you run. Vet them as you would any dependency.
  • General purpose means broad reach. Decide what the agent may touch outside a repository before it touches it.

How it scores on openness

  • Replaceable. Strong. Many providers and local models.
  • Inspectable. Strong. Apache 2.0 source under foundation governance.
  • Portable. Strong. Extensions are MCP servers and recipes are YAML files.
  • Bounded. Depends on configuration: which extensions are enabled defines what it can do.
  • Grounded. Good when extensions connect it to real systems of record.
  • Auditable. Achievable. Local operation means you control what is recorded.

What it is not

goose is not a model and not tied to one provider. It is not a broker; it runs its own agent. It is not only a coding tool.

Openness scorecard for goose (AI agent)

The notes above describe the tool. The score belongs to your deployment of it, because configuration, permissions, and where records are kept decide most of the result. Score each test 0 (absent), 1 (partial), or 2 (demonstrated) for the way you run goose (AI agent), out of 12.

TestQuestionWhat a 2 looks likeEvidence to collect
ReplaceableCan one component be swapped without rebuilding the system?A major component has been swapped, and the work was measured in days, not months.Which component did you last swap, and how long did it take?
InspectableCan a builder understand what runs and why?The exact context, tool calls, and results for each step are exported to storage you own.Where do traces live, what do they contain, and how long are they kept?
PortableCan identity, skills, context, and work move?Identity, skills, and work are files in open formats and have been loaded into a second tool.What have you exported and reloaded somewhere else?
BoundedAre authority and approval requirements explicit?Tool scopes, delegated user identity, and approvals are enforced outside the model and tested adversarially.What stops the agent from taking an action its user could not take?
GroundedDo agents share durable data and semantic meaning?Agents query governed metrics over versioned tables and cite the definition and data version.Which semantic definitions and tables do agents use?
AuditableCan people reconstruct decisions and outcomes?Runs, approvals, and data versions are recorded long enough to reconstruct any decision.Could you reconstruct an agent decision from three months ago?

Score your goose (AI agent) setup on the interactive scorecard. It runs in your browser, and you can download the result as JSON or print it.

Where to learn more

Primary sources first. Documentation and specifications move faster than any summary, so treat the links below as the authority and this page as orientation.

Work with Alex