Crewlet Operator Documentation
The Crewlet Operator runs Crewlet AI agent companies on Kubernetes: declare a company and its org chart as custom resources, and the operator runs its engine.
You write the Company, its units and seats, its model providers, tool
servers, integrations and schedules as crewlet.ai objects. The operator
renders them into the one company document Crewlet
runs on and applies it whole or not at all, pushes the credentials they
reference into the engine’s encrypted store, and runs the engine’s pods,
routes and network policy beside them.
New here? Read in this order: Quickstart → Installing the Operator → How the Operator Models a Company, then the guide for what you are setting up.
Caution: alpha. The crewlet.ai/v1alpha1 API changes without notice.
Each version of these pages describes the operator it was published with.
Getting Started
Section titled “Getting Started”Run the operator and a two-agent company on a local kind cluster: change it, break it on purpose, open its dashboard and remove it all again
A production install with the Helm chart: prerequisites, the engine image to run, the values that matter, and the roles and admission policies it adds
Core Concepts
Section titled “Core Concepts”How the operator turns objects into a running engine, one part per page:
The eight kinds and how they relate, the identities that never change, credentials by reference, management policies and what every object reports
The company document and each engine pod’s crewlet.yaml the operator renders from the objects, where every seat is placed, and what blocks a document
What happens on every change: coalescing, the secret store, the dry run and the apply, management policies and drift, seat pauses, and every condition an object reports
Every object the operator creates for a Company: the data pod, seat pools, exposure and network policy, versions, suspension, deletion and backups
Every position in a Crewlet company that holds a credential and the field that fills it, who can read a value, and how the engine’s store entries are named
Guides
Section titled “Guides”Write the Secret for each credential, rotate it without a restart, and find out why one did not arrive
Slack, GitHub, GitLab, Mattermost, Jira, Confluence, Datadog and an Atlassian organization: each Integration, the seats’ own credentials and the accounts the engine creates
MCPServer objects: an engine image with each stdio server installed, the server each integration’s agents use, and how to tell that one did not start
CLIAgent providers: an engine image with the CLI, a subscription login in a pod, text mode and agent mode, and the provider a seat’s code work runs on
The public listener and the webhook routes through a Gateway, an Ingress or your own proxy, the internal API route, the dashboard’s address and the policy around each engine
A Human seat for each person, where people sign in, the two-step verification a class may make optional, the company’s first person and everyone after, and why the operator’s tokens hold no seat
Enforce, Observe and Seed: who owns the company document, which to choose, and how to move a running engine under the operator
Data pods and stateless seat pools, Shared and PerSeat, what a pod of its own isolates and what it costs, and where each seat runs now
A control plane that provisions tenants: a namespace each, server-side apply, the status to read, credentials as Secrets, tiers as classes, offboarding
Every condition and reason an object reports, what each one means, and what to do about it
Reference
Section titled “Reference”Every kind and field of crewlet.ai/v1alpha1, generated from the API types with make api-docs
For working on the operator itself, see CONTRIBUTING.md; for publishing it, RELEASING.md.
Part of Crewlet. Generated from crewlet/crewlet-operator main at d1515a2.