Current Progress
Spillway is in a private TestFlight beta.
This page gives a broad product-level snapshot of what is usable today and what remains emerging. The deeper architecture documents govern how Spillway is designed; this page intentionally avoids internal issue numbers, implementation slices, test counts, and release-engineering detail.
Available in useful form
Mac
The Mac app supports everyday Gmail reading and organization, multiple accounts, composition, attachments, user-controlled categories, review and correction workflows, local and optional cloud AI, and recovery tooling.
It is currently the most mature daily-work surface.
iPhone
The iPhone app supports ordinary mail use, triage and action workflows, attachments, and selected cross-device continuity.
It is useful on its own, while some deeper Work, learning, and administration workflows remain more mature on Mac.
Categories, attention, and organization
Spillway supports user-controlled categories, multiple categories per message, review and correction, and category evolution without treating past decisions as disposable.
Its attention views are designed to reduce scanning while keeping human choices distinct from machine suggestions.
AI, evidence, and teaching
The app can use local AI or explicitly configured cloud AI providers.
Spillway preserves important distinctions among source facts, machine inferences, and human decisions so later reasoning can retain provenance rather than flattening everything into one score.
On Mac, Teach Spillway lets a person record natural-language guidance. That guidance can inform later prioritization without automatically becoming permission to file, hide, delete, or otherwise act on mail.
Cross-device continuity
Selected product state can move between supported Apple devices.
Synchronization is designed to preserve what information means and where it came from. A machine suggestion does not become a human decision merely because it moved to another device.
Knowledge and background intelligence
Spillway has working infrastructure for selecting bounded relevant context and for performing background analysis without treating idle compute as something that must be consumed.
The architecture is designed for accumulated evidence and knowledge to improve later decisions selectively rather than requiring every message to be reconsidered from scratch.
Emerging
Several product areas are usable in partial form and continue to grow:
- richer Work views for requests, commitments, waiting states, dates, and opportunities;
- broader use of accumulated knowledge and relationships to improve later inference;
- empirical calibration from human decisions and outcomes;
- deeper cross-device continuity for learning and Work;
- more complete historical learning with honest coverage reporting;
- stronger system handoff to calendars, sharing, files, and other apps.
These are parts of Spillway's intended architecture, not promises that every described capability is available in the current beta.
Planned direction
Spillway is intended to expand beyond Gmail and beyond email while preserving source identity, privacy boundaries, provenance, and user control.
Future sources may include other mail systems, calendars, documents, chat systems, and related user-approved inputs.
The architecture separates the source of information, the kind of artifact, the AI or deterministic process interpreting it, and the external system receiving a handoff. That is intended to let Spillway grow without rebuilding its evidence and knowledge model for each new integration.
Important limits
Spillway is still beta software.
Gmail is the supported mail provider today. Some learning and Work capabilities remain Mac-local or incomplete across devices. Features may change, and testers should keep access to Gmail or another established email client rather than relying on Spillway as the only way to reach important mail.
For testing expectations and participation details, see Private Beta.
Explore features · How Spillway is built · Private beta
Documentation provenance: Iterative human–AI construction. See Documentation Provenance.