Model tasks that work like commands¶
Ribbit is an extensible framework for composing model tasks into reusable commands, with typed contracts and inspectable evidence. It reads UTF-8 text, JSON values, JSONL, files, and tool output. Keep the evidence, choose a route for each semantic step, and pass the result to another command or a coding harness.
These docs describe the source checkout. The latest published preview is 0.1.0-alpha.2; the unreleased changelog lists changes not yet in the npm alpha or published native archives.
New to Ribbit? Start with your first pipeline. It needs no model, account, or provider configuration.
| What you need | Where to go |
|---|---|
| Learn by building something small | Tutorials |
| How do I do...? | How-to guides |
| Look up syntax, formats, or limits | Reference |
| Understand Ribbit's core | Explanation |
Try a workflow¶
- Prioritize customer feedback: classify locally, then synthesize with a stronger model.
- Prepare repository context for a harness: retain source text and attach relevance labels.
- Investigate a failing CI check: interpret a controlled offline failure from saved evidence.
- Prepare a date-fns contribution brief: pin repository sources, then annotate a brief locally.
- Analyze an open dataset: normalize, select, and annotate a small deterministic sample.
- Save a useful transformation: turn a command into a named definition and a flow.
Install from npm or a native GitHub archive. See installation for prerequisites and platform support.
Ribbit is an MIT-licensed alpha. Semantic results depend on the model; the current evaluations show both passing and failing cases. Linux, macOS, and Windows have native x64/ARM64 builds. Interactive Windows picking remains separately unverified.
These docs use Diátaxis: lessons, task guides, technical reference, and explanations are separate so you can find the kind of help you need.