goose

Why it's ranked here

Governance and protocol discipline are the argument here. The project is Apache-2.0 licensed, and its README places it under the Agentic AI Foundation at the Linux Foundation, a stronger continuity signal than a single-vendor repository. The model and the toolset swap independently, both through published protocols rather than bespoke adapters, and local inference sits inside the provider layer rather than beside it. HTTP-level test scaffolding is wired into the build, though this read could not establish how far that coverage actually reaches.

What's good

The agent-safety controls are real flags, not documentation: you can cap total turns, cap consecutive identical tool calls, and run extensions inside a Docker container. Local execution is genuinely local, with two inference backends, one general-purpose and one for Apple silicon, behind a cache that loads, evicts and recommends models based on available memory, so it runs with no hosted provider at all. The published container image builds only the command-line package, pins its runtime base by digest, and drops to a non-root user. Even unusual build settings are explained in comments.

Tradeoffs

This is a large surface to depend on: the core library exposes more than forty public modules, covering everything from scheduling to dictation. The build is heavy and toolchain-sensitive, with a vendored JavaScript engine, one dependency pulled from a pinned git revision, and another exact-pinned next to a note that it ships breaking changes in patch releases. The shipped container image also builds with an older Rust toolchain than the workspace declares it requires. Documentation points entirely at an external site, and the repository badges name a different GitHub organisation than the catalogue entry links to.

How to use it well

Treat the guardrails as day-one configuration: cap turns and repeated tool calls before pointing it at anything that costs money or writes to a real repository. The container option runs extensions inside a Docker image you name, but the extension must already exist in that image, so build it before relying on it. Move repeatable prompts into recipes and pass values as parameters; built-in commands show what a recipe expands to before you run it. When a run must be reproducible, pin the extension set to exactly what you pass on the command line, and for scripting, request structured JSON events instead of prose.

Observed

License

Apache-2.0, declared in the workspace manifest

Language and packaging

Rust, built as a multi-crate Cargo workspace

Interfaces

Terminal CLI, native desktop app, and a JSON-RPC server speaking the Agent Client Protocol

Local inference

Two local backends, llama.cpp and MLX, so no hosted provider is required

Core library surface

Over forty public modules declared in the core crate