Quickstart
Start here
Documentation home Quickstart Glossary Changelog
In this guide
Prerequisites Install the CLI Create a project Attach knowledge Validate and test Run a case Read the evidence Approve the task Audit and credits
Target: a governed run finished, evidenced, approved and audited in under twenty minutes. We use the Content Studio Crew Template because it is the lowest-risk template, needs no connectors and runs on a sample corpus we ship with it. Everything you learn here applies unchanged to LawCrew, Finance Close and AquaMind.
You will spend roughly 90 credits. Your trial has 5,000, so this costs about 1.8% of it. Credit consumption is shown at every step because surprise billing is the fastest way to lose a technical evaluator.
1. Prerequisites
- A KrewOS account — start free, no card, 5,000 credits
- Node 20 or later, or any environment that can run a single static binary
- About twenty minutes and a terminal
You do not need model provider keys. The trial routes inference through KrewOS-managed keys. If you want to use your own from the start, skip ahead to configuring BYOK and come back.
2. Install and authenticate the CLI
# install
npm install -g @krewos/cli
# authenticate against your workspace — opens a browser, stores a scoped token
krewos auth login --region sg
✓ Signed in as [email protected]
✓ Tenant ridgeway-partners Region sg-central-1
✓ Wallet 5,000 credits Expires in 30 days
The CLI and the Studio are two views of the same manifest. Anything you do here shows up in the browser immediately, and vice versa.
3. Create a project from a Crew Template
Template-first is the default. A blank canvas is available but de-emphasised, because starting from something that already passes an evaluation suite is faster than starting from nothing.
# list the templates available to your tenant
krewos template list
content-studio Content Studio v4.2.0 risk: low
lawcrew-doc-review LawCrew Document Review v3.0.2 risk: high · pack required
finance-close Finance Close v2.4.1 risk: high · pack required
aquamind-advisory AquaMind Advisory v2.7.1 risk: medium · pack required
bizdev-intel BizDev Intelligence v1.9.4 risk: low
ecom-selection eCommerce Selection v1.3.0 risk: low
# scaffold a project in the development environment
krewos project create quickstart \
--template content-studio \
--language en,zh-Hans \
--jurisdiction SG \
--env development
✓ Project quickstart created
✓ Manifest written to ./quickstart/crew.manifest.yaml
✓ 1 Crew · 3 Pods · 9 Agents · 2 human wait states · 1 declared Action
What just got written
The manifest is the source of truth. The canvas is an editor over it, plain-English changes compile into it, and Git commits resolve to it.
krewos manifest show --summary
crew: content-studio
evidence_floor: 0.72 # abstain below this sufficiency score
autonomy: full_review # every crew starts here, always
effects:
- kind: publish_cms
idempotency: required
approval: senior_editor # gate proven at compile time
knowledge:
- pack: sample-corpus-en # shipped with the template
4. Attach your own knowledge (optional)
The sample corpus is enough to finish this guide. If you would rather see retrieval over your own material, build a Knowledge Pack now — it takes about two minutes for a small folder.
krewos knowledge create house-style \
--from ./docs \
--acl workspace \
--freshness 180d
✓ Indexed 34 documents · 1,182 chunks · 41 credits
# attach it to the crew and re-resolve the manifest
krewos project attach-knowledge house-style
Freshness windows matter. When a source expires or materially changes, every artefact that relied on it is identified and its owners notified. A Knowledge Pack with no freshness window will not raise that signal.
5. Validate and test before you run
Validation is static. It proves properties about the crew rather than testing them — including that no path runs from untrusted input to an external side effect without an intervening approval gate.
krewos validate
✓ Typed ports all 14 connections resolve
✓ No orphan nodes 0 found
✓ Bounded loops revision loop capped at 3 iterations
✓ Effect declarations 1 declared, 1 gated
✓ Taint analysis no untrusted input reaches publish_cms without approval
✓ Permissions no node holds more scope than it declares
✓ Manifest valid ready to test
krewos test --sample
✓ 6 sample cases · 6 passed · evidence floor respected on all
cost: 34 credits (platform 12 · model 22)
If validation fails, there is no release to deploy. That is the point — the compiler is a gate, not a linter.
6. Run a real case
krewos run start \
--input ./samples/brief-sustainable-aquafeed.json \
--watch
✓ RUN-9C41A8 admitted · reserved 260 credits
[00:04] Research Pod 7 retrievals · sufficiency 0.86
[00:31] Outline accepted
[01:12] Drafting 2,140 words · 11 citations
[01:58] Compliance gate passed · 0 flags
[02:10] Revision loop 1 of 3 · readability
[02:44] Senior editor gate waiting for a human
run parked at CP-6 · this is a normal state, not a timeout
The run is now durable. It will survive a worker restart, a deploy and a rollback, and it will wait days for the approval without consuming anything but a few credits of task management.
7. Read the evidence behind a claim
This is the step that matters. Open the run in Studio, or inspect it from the terminal.
krewos run evidence RUN-9C41A8 --claim c-07
claim "Feed conversion ratios below 1.2 are achievable in
recirculating systems with stable dissolved oxygen."
sufficiency 0.86 floor 0.72 · above floor, no abstention
sources
[1] SFA Aquaculture Advisory 2024/03 authority: regulator · dated 2024-08-14
[2] Tan & Loh, Aquaculture Research authority: peer-reviewed · dated 2023-11-02
[3] Internal trial log, Pulau Ubin site authority: first-party · dated 2025-02-19
conflicts none
staleness source [2] expires in 214 days
Try forcing an abstention: run the same crew with the knowledge pack detached. The crew will not guess. It returns a structured insufficiency result naming what it could not establish and which sources would resolve it, and the case routes to the knowledge-gap queue rather than the failure log.
8. Approve the human task
Approvals are durable workflow states, not emails. The reviewer sees the business purpose, the requested decision, the risk classification, the diff, the cited sources with authority and dates, any failed or marginal gates, and what runs next after approval.
krewos task list --mine
TASK-4471 Senior editor gate RUN-9C41A8 waiting 6m risk: low
krewos task approve TASK-4471 --note "Citations check out, publish."
✓ Correction Record CR-1182 written · reviewer s.rajah · immutable
✓ Action publish_cms fired with idempotency key run-9C41A8-cp7
✓ RUN-9C41A8 complete
9. Read the audit trail and the credits
krewos run audit RUN-9C41A8 --format json > audit.json
✓ 41 trace records · 9 nodes · 14 model calls · 7 retrievals · 1 human decision
krewos run cost RUN-9C41A8
platform 52 credits S$0.52
model 38 credits S$0.38
reserved 260 released 170 at close
settled 90 credits S$0.90
outcome unit 1 article published
cost/outcome S$0.90
The audit export contains the run, the release, every component version, the model profiles, the evidence with provenance, the human decision with identity and timestamp, the policy outcomes and the cost records. It is the artefact you hand an auditor.
Where to go next
- Engineers — export to Git and extend with a custom Tool, or read the Crew Manifest reference
- Compliance — evaluation suites and release gates, then the Trust Centre
- Buyers — put your real volume into the estimator and compare it to the ledger you just read
On this page
1 · Prerequisites 2 · Install the CLI 3 · Create a project 4 · Attach knowledge 5 · Validate and test 6 · Run a case 7 · Read the evidence 8 · Approve the task 9 · Audit and credits Where to go next