KrewOS documentation

Start here

Documentation home Quickstart Glossary Changelog

Everything you need to build, govern, operate and extend AI Agent Crews on KrewOS. If you have twenty minutes, start with the Quickstart — it ends with you reading a real evidence record and approving a real human task.

Four paths

Pick the one that matches what you are trying to do today. They interlink heavily, so you will not be stuck in a lane.

Build

Assemble a crew from verified components without writing orchestration code, then take it to Git when you want to.

• Quickstart — a governed run in under twenty minutes
• Your first crew from a Crew Template — configure jurisdiction, language and risk level
• Adding knowledge — build a Knowledge Pack from your own documents, with access control and freshness windows

Govern

The mechanics that are enforced by the compiler and the runtime, outside the model prompt, where a prompt injection cannot reach them.

• Evidence and abstention — sufficiency scores, declared evidence floors, and what a structured insufficiency result contains
• Autonomy policies — the four stages, entry conditions, review rates and automatic reversion triggers
• Release gates — evaluation thresholds that block promotion, and how to set one that is strict without being unshippable

Operate

Running crews in production, where things wait days for a human and must never fire the same external Action twice.

• Runs and recovery — durable state, checkpoints, idempotency keys and compensation behaviour
• The approval queue — task routing, SLAs, escalation and dual approval
• Cost and outcomes — credits, budgets, Outcome Units and cost per outcome against your baseline

Extend

The engineering surface. Manifest-first, Git-native, and ejectable at any time.

• CLI — krewos validate, krewos test, krewos package, krewos deploy
• Custom Tools and Actions — the published extension contract and the sandbox rules
• Exporting to Git — linked projects that round-trip through a reviewed diff, and ejected projects that are yours outright

Reference

The canonical technical artefacts. The Crew Manifest schema is published as a public contract — you can validate against it without an account.

• Crew Manifest reference — every field, with a complete worked example
• API reference — eleven domains, authentication, events, webhooks, errors, rate limits, OpenAPI download
• CLI reference — install, authenticate, validate, test, package, deploy, inspect
• SDK reference — language coverage and the component contract

Tutorials

Eight end-to-end guides, each one a job somebody actually has to do.

1. Build a document review crew from the LawCrew template
2. Create a Knowledge Pack from your own documents, with ACLs and freshness windows
3. Configure BYOK and a model policy with a fallback chain
4. Write an evaluation suite and set a release gate
5. Add a governed Action through n8n with an idempotency key
6. Expose your crew as an MCP tool to an external assistant
7. Export to Git and extend with a custom Tool
8. Graduate a crew from full review to supervised sampling

Conventions in these docs

Vocabulary is shared with the product and the glossary renders from the same source, so a term on this page means exactly what it means on the screen you will see later.

• Tools read or compute. Actions cause external side effects. The distinction is load-bearing and we never blur it.
• Knowledge is not Memory. Knowledge is source-backed organisational content retrieved at run time; Memory is task state.
• Runtime capitalised means the KrewOS Runtime surface; lower-case "execution" is the generic sense.
• British spelling throughout — organisation, licence as a noun, catalogue, programme.

Getting help

Search covers docs, the glossary, the Manifest schema and the changelog together. If the answer is not there, the support address on your plan is in the workspace footer, and security questions go to [email protected] regardless of plan.

On this page

Four paths Build Govern Operate Extend Reference Tutorials Conventions Getting help