Documentation

SayeOS documentation.

How SayeOS Core works: getting started, the core concepts, the agnostic architecture, building with Skills and Python, and the governance behind it.

Areas

Getting started

Start here.

What is SayeOS Core?

A free, governed workspace for persistent intelligent work. Core is model, system, and connection agnostic, and public distribution is coming soon.

Preview access

Core is not yet publicly distributed. Join the preview from the homepage for release updates and early access.

Supported environment

As of July 2026, supported environment details are being finalized. Preview members receive environment updates first.

Models and providers

Connect the model or token provider you prefer. You connect and pay your provider separately where applicable.

Workstreams

Work in Core happens in persistent workstreams that carry intent, context, state, evidence, decisions, and history.

Core concepts

The primitives.

Workstream

A persistent unit of work that continues across sessions and contributors, holding everything the work depends on.

Intent

The human originated purpose that authorizes work. Intent is declared, not inferred.

Context

What matters for this work: the understanding the workstream carries.

State

What is currently true in the work right now.

Evidence

What supports a conclusion or action, sufficient to reconstruct it later.

Review

The point where a human checks work before it counts. Worker output is draft until reviewed.

Skills

Bounded, executable capabilities with defined inputs, allowed tools, permitted actions, and review.

Continuum

How understanding persists and evolves over time while keeping source, context, confidence, and sharing boundaries. See the Continuum page.

Workstream ContextWhat matters forthis work StateWhat is currentlytrue EvidenceWhat supports theconclusion
Context, State, and Evidence are three separate primitives, each connected to the workstream.

Architecture

Agnostic by design.

System agnostic

SayeOS coordinates work around existing systems of record rather than replacing them. Those systems keep their authority.

Model agnostic

Different models contribute by task, cost, capability, privacy, environment, and policy. Changing providers does not restart the work.

Connection agnostic

Connections include APIs, local tools, files, databases, approved services, Edge mediated access, and defined interfaces. No one method is universal.

Python execution

Python is a first class execution layer. Python workers run defined analysis inside the workstream, bounded by security, dependencies, and permissions.

Systems of record

Operational systems remain the authoritative source for their own data. SayeOS references them rather than absorbing them.

Workers and contributors

People, models, Python workers, APIs, software services, and reviewers can all contribute to one workstream.

Build

Extend Core.

Models reason. Python executes. Skills and Context Packs make capability reusable. See the Build page.

Skills

Define inputs, required context, allowed tools, permissions, validation, and review, then test, package, and publish.

Context Packs

Package governed knowledge with known provenance and scope for reuse.

Python

Use appropriate libraries for the domain. Execution is bounded, not unrestricted.

Packaging

Bundle a capability with its metadata: version, permissions, dependencies, and limitations.

Validation

Define the checks a Skill must pass before its output is trusted.

Governance

The rules behind the work.

SayeOS runs under a public Constitution. Read the governing text.

Why a Constitution?

Consequential work should run under rules you can read. The Constitution is public and versioned.

Human authority

Human authority stays primary. Intelligence may analyze, simulate, and propose, but may not decide or approve.

Pre-authorization

A person can delegate scoped authority in advance. Systems execute inside that scope while it is valid and not revoked.

Refusal

When an instruction is ambiguous or would require guessing intent, halting and escalating is the correct behavior.

Escalation

Escalations surface to a human authority who can resolve them. They are not resolved by retries or heuristics.

Questions?

Read the FAQ.