For engineering
Manifest-first. Git-native. Ejectable.
If you are the person who will be blamed when this is unmaintainable in eighteen months, this page is for you. The short version: the canonical artefact is a file, it lives in your repository if you want it to, and you can leave with everything.
You are not trapped.
Export a complete engineering repository at any time. Not a JSON blob — a project that builds, tests and deploys.
$ krewos export --project lawcrew-doc-review --mode eject
lawcrew-doc-review/
crew-manifest.yaml # canonical source of truth
components/ # agents, pods, deterministic nodes
schemas/ # typed port contracts
prompts/ # versioned, with bindings
policies/ # deterministic rules
evaluations/ # suites, datasets, scorers
tests/ # unit + integration
extensions/ # your custom tools and actions
deploy/ # container + IaC
AGENTS.md CLAUDE.md # agent instructions
README.md
ejected this project is yours. runtime APIs continue to work.
Linked projects round-trip
Keep the project linked and changes flow both ways through a reviewed diff. A Studio edit appears as a proposed commit; a commit appears as a change plan a reviewer accepts or rejects. No silent overwrite in either direction.
Ejected projects are yours
Full stop. No licence revocation clause on your manifests, no phone-home requirement, and the deployed runtime APIs keep working. Packs are separately licensed and that licence is what ends — not your code.
Everything is scriptable
Validate, test, package, deploy and inspect from the CLI. The web Studio calls the same API surface you do; there is no privileged path.
Eleven API domains
Runs, approvals, releases, components, knowledge, evaluations, audit, cost, secrets, identity and webhooks. OpenAPI spec published, with a downloadable collection.
The manifest schema is published and versioned. If you want to generate manifests from your own tooling and never open Studio, that is a supported way to use the product, not a workaround.
What we own so you do not.
Each of these is a system you would otherwise build, staff and carry. The full build-versus-buy breakdown →
Durable state and checkpointing
Runs that survive worker restarts and deploys, with replay semantics you do not have to reason about.
Idempotency and compensation
Keys, retries and declared compensating actions on every externally visible effect.
Secret injection
Credentials injected at call time by the gateway, never present in prompts, logs, manifests or exports.
Permission enforcement outside the prompt
A component cannot talk its way into a permission it was not granted, because the grant is not in the context window.
Cost attribution
Per-node metering that reconciles to a business outcome unit with no unattributed spend.
Evaluation harness and audit trail
Datasets, scorers, confidence bounds, release gating, and an immutable exportable trace.
Extension contract and sandbox rules.
Stated plainly, because engineers respect a clear boundary more than a generous one.
| You can write | Against | Appears in Studio as |
|---|---|---|
| Custom Tools | Published Tool interface — typed input and output, no external side effect. | A block, with its contract shown on the node |
| Custom Actions | Action interface — must supply an idempotency key and declare compensation behaviour. | A block flagged as effect-causing |
| Deterministic nodes | Pure function interface. No model call, fully replayable. | A block the compiler can reason about |
| Scorers | Evaluation scorer interface — case in, score and rationale out. | An option in the Evaluation Lab |
| Connectors | Connector interface with a declared permission scope and read/write classification. | An entry in your Tools registry |
Sandboxed execution
Generated and custom code runs sandboxed with network and filesystem restrictions declared per component. Nothing runs with ambient credentials.
Dependency scanning
Every dependency is scanned on registration and on a schedule. A new critical advisory on a registered component notifies its owner and blocks new releases that include it.
Review before registration
Custom components are reviewed before they are registered in a tenant catalogue, by your reviewers under your bar. The workflow is the same one the Hub uses.
Quickstart, on this page.
Twenty minutes from nothing to a governed run with a readable evidence record.
# 1. install and authenticate
$ npm i -g @krewos/cli
$ krewos auth login # opens browser, stores a scoped token
# 2. start from a template rather than a blank canvas
$ krewos init --template content-studio --name first-crew
resolved [email protected] · 11 components · pinned
# 3. attach the sample corpus and validate
$ krewos knowledge add ./samples/corpus --pack sample-corpus
$ krewos validate
schema ok · ports ok · effects ok · budget ok
# 4. run a sample case in the development environment
$ krewos run --case ./samples/brief-01.json --watch
research done 8 sources, sufficiency 0.81 (floor 0.70)
draft done 1,140 words, 8 claims cited
compliance done 6 checks passed
approval waiting task assigned to you
# 5. approve, then read the trace
$ krewos approve --task t-7741 --note "tone fine, sources check out"
$ krewos trace --run r-8f21 --format json > trace.json
42 events · cost 38 credits (S$0.38) · 0 unattributed