id: repo:kdcube/app/ai-app/src/kdcube-ai-app/kdcube_cli/README.md title: "KDCube CLI" summary: "Installs and operates KDCube runtimes, checks the active app-owned delegated catalog, manages exact secrets through their selected provider, and exposes typed deployment and management APIs for application-specific CLIs." tags: ["kdcube", "cli", "runtime", "deployment-target", "secrets", "python-api"] keywords: ["kdcube-cli", "kdcube init", "kdcube start", "kdcube stop", "kdcube secrets", "kdcube bundle catalog check", "app-owned delegated catalog", "kdcube bundle delete", "managed bundle deletion", "purge-data", "force-retire", "targeted bundle retirement", "kdcube_cli.control", "kdcube_cli.management"] updated_at: 2026-09-13 see_also:

• repo:kdcube/app/ai-app/docs/service/cicd/cli-README.md
• repo:kdcube/app/ai-app/docs/service/cicd/deployment-target-control-api-README.md
• repo:kdcube-ai-app/app/ai-app/docs/service/secrets/secret-management-cli-README.md
• repo:kdcube/app/ai-app/docs/service/secrets/host-vault-README.md

KDCube CLI

Bootstrap and operate a KDCube platform stack from the command line.

Install

pip install kdcube-cli

Or with pipx (recommended):

pipx install kdcube-cli

Python control API

Application-specific CLIs can import kdcube_cli.control instead of importing private command helpers or launching the kdcube executable. The typed API selects local or endpoint-only targets, reports capabilities, controls the supported local lifecycle, and resolves an installed app's public surfaces.

from pathlib import Path

from kdcube_cli.control import ApplicationRef, LocalDeploymentTarget, select_local_target

reference = select_local_target(Path("~/.kdcube/kdcube-runtime/acme__prod"))
target = LocalDeploymentTarget(reference)
url = target.application_url(ApplicationRef("connection-hub@1-0"))

See the deployment-target control API for capabilities, explicit surface selection, endpoint-only behavior, and structured errors.

Secret Management

Logical secret operations are backend-neutral. They address one exact platform or bundle key through the running KDCube management service, which uses the ISecretsManager selected by that deployment:

kdcube secrets metadata platform.services.brave.api_key --scope platform
kdcube secrets set platform.services.brave.api_key --scope platform
kdcube secrets get platform.services.brave.api_key --scope platform --output ./secret.txt
kdcube secrets delete platform.services.brave.api_key --scope platform

Delegated operations require an exact live Card grant and accept the bearer only through a hidden prompt or --credential-stdin. get writes a private file while the terminal receives a receipt. export is a separate browser-confirmed, one-use owner ceremony that reconstructs only explicitly named descriptor keys.

Storage selection and migration are grouped beneath secrets backend:

kdcube secrets backend status --tenant <tenant> --project <project> --json
kdcube secrets backend host-vault prepare --tenant <tenant> --project <project>
kdcube secrets backend host-vault stage --tenant <tenant> --project <project>
kdcube secrets backend host-vault activate --tenant <tenant> --project <project> --yes

The former kdcube secrets host-vault ... spelling remains an alias. Read Manage KDCube Secrets for local and remote targeting, human and agent credential input, consent recovery, exact-key operations, export, and backend status. Read Host Vault for Provider Secrets for provisioning, mTLS identity, migration, activation, and recovery.

When secrets.service.backend is host-vault, the vault is a startup dependency: chat-ingress and chat-proc wait for the kdcube-secrets health check, which fails while the vault is unreachable. kdcube start, kdcube refresh and kdcube init therefore check the vault before Compose runs. They start it only when the descriptor declares secrets.service.host_vault.local_service (macOS, and Linux with an explicit bind, and refused on Windows). With any other backend none of this runs.

What You Build

With KDCube, you author portable application bundles.

A bundle can be a full application, an internal tool, a workflow backend, a UI-backed product surface, an MCP server, a scheduled automation, an agent, or a mix of these. It can use KDCube's agent facilities when agentic behavior is needed, or it can be ordinary application code.

You focus on product behavior:

• what the app or workflow should do
• which APIs, screens, jobs, tools, MCP servers, or webhooks it exposes
• what state it reads and writes
• how users interact through chat, UI, messages, or external events
• which optional agent capabilities or framework adapters it uses, if any
• how many conversations, tasks, or long-running threads it maintains

KDCube provides the hosting runtime around it: tenant/project isolation, auth, routing, streaming, storage, conversation/message handling, service discovery, hot reload, deployment wiring, and reusable application services.

When an app needs an agent workflow, its optional agent capabilities can include the native KDCube ReAct Agent, adapters for framework-owned runtimes such as LangGraph and Claude Code, configurable tool and skill execution, isolated code execution, MCP access, streaming progress, artifacts, and memory/search facilities. KDCube calls this adapter path the hosted foreign-runtime path; the host layer is runtime-agnostic and each adapter binds the shared harness contracts its agent needs. An app can also run Claude Code directly for app-owned work without binding it to the conversational harness.

In other words: you author the application module; KDCube hosts it and gives it access to the platform services and optional agent facilities it needs.

What is a bundle?

A bundle is the deployable application unit of the KDCube platform.

Concretely, it is a folder or git-backed source reference that contains bundle code plus metadata describing the surfaces KDCube should expose. The platform discovers the bundle, loads its entrypoint, wires its declared surfaces into the runtime, and manages reload/lifecycle for it.

A bundle can expose any combination of:

• HTTP APIs — authenticated operations APIs or public webhook endpoints
• Frontend assets — bundle-owned UI/static assets served by the platform
• MCP servers — Model Context Protocol endpoints for agent/tool use
• Scheduled jobs — cron-driven background automation
• Message handlers — conversation/message workflows with attachments, external events, steer, and followup
• Agent workflows: the native ReAct Agent plus LangGraph, Claude Code, and other framework-owned agent runtimes hosted through KDCube adapters, each with the tools, skills, MCP connections, and code-execution facilities its adapter binds

Tenant, project, and workdir

Every KDCube runtime lives in a namespaced workdir:

~/.kdcube/kdcube-runtime/<tenant>__<project>/

tenant and project together define one isolated environment — its own config, data, credentials, Postgres/Redis stores, and running stack. Use separate namespaces for separate customers, products, or lifecycle stages (dev, staging, prod).

~/.kdcube/kdcube-runtime/
├── default__default/       # default scope
├── acme__staging/          # acme tenant, staging project
└── acme__prod/             # acme tenant, prod project

Each scope is fully isolated — its own config, data, logs, and running stack.

For local seed descriptors, storage and host path fields may be left null or omitted to use the tenant/project workdir defaults. For example, storage.kdcube: null and storage.bundles: null resolve to:

~/.kdcube/kdcube-runtime/<tenant>__<project>/data/kdcube-storage
~/.kdcube/kdcube-runtime/<tenant>__<project>/data/bundle-storage

Explicit file:///... values select a custom local host path. Explicit s3://... values are preserved as remote storage URIs.

One machine, one running stack

A machine can hold many workdirs on disk, but only one stack can run at a time. Starting a second workdir while another is live aborts with a message showing what is running and how to stop it first.

One workdir, many bundles

Inside one tenant/project environment you can register and run any number of bundles. They share the same platform infrastructure — storage, auth, Postgres, Redis, and the same deployment boundary.

This is the normal model: one environment, multiple application modules running side by side.

Bundles are portable across workdirs

A bundle is just code (a local path or a git repo) plus a descriptor entry in bundles.yaml. The same bundle can be registered in multiple workdirs independently — each workdir maintains its own config, secrets, and runtime state for that bundle. This makes it straightforward to promote a bundle from a dev environment to staging or prod by registering it in the target workdir and supplying the appropriate descriptor values.

Get started

init is for first-time setup only

kdcube init creates a brand-new namespaced runtime workdir. It refuses if the target workdir is already initialized (i.e. has install-meta.json). To pick up platform code changes or rebuild images on an existing workdir, use kdcube refresh instead.

Interactive first run

The fastest way to get a local KDCube stack running is to pick a tenant and a project. In a terminal, init opens the first-run prompts and creates the runtime under the platform default base ~/.kdcube/kdcube-runtime/<tenant>__<project>/:

kdcube init --tenant acme --project staging
kdcube start --tenant acme --project staging

With no source selector, init uses the latest published platform release. It offers Google login by default, then optional model-provider and private-Git credentials. Bare kdcube init also works in a terminal: it first asks for the tenant and project. Piped and CI invocations do not prompt; use --non-interactive and supply all required inputs as flags or descriptors.

The default init stages the configured base set: Connection Hub (identity, consent, delegated credentials), KDCube Services (managed MCP + named services), User Memories, and the workspace showcase app — as a pure config overlay on the bundles shipped in the image. Default identity uses server-side Google login: the workspace app validates Google identity and Connection Hub issues the KDCube session.

Init ends with a first-run checklist of placeholders still unfilled; features backed by an unfilled slot stay inactive, everything else runs. Step-by-step recipes: clean install · from a descriptor set · daily operations.

To request the standard secret prompts explicitly:

kdcube init --tenant acme --project staging --prompt-secrets

To stage known secret values without prompts, use dotted descriptor keys:

kdcube init --tenant acme --project staging \
  --set-secret platform.services.openai.api_key "sk-..." \
  --set-secret platform.services.anthropic.api_key "sk-ant-..."

Descriptor-driven init (reproducible / automated)

When you have a descriptor set (assembly.yaml, bundles.yaml, etc.):

kdcube init --tenant acme --project staging \
  --descriptors-location /path/to/descriptors

With a local platform source tree and image build:

kdcube init --tenant acme --project staging \
  --descriptors-location /path/to/descriptors \
  --path /path/to/kdcube-ai-app \
  --build

Authentication

The default identity is server-side Google login defined by the Workspace app (the CLI value is bundle). Select a method in the prompt or pass --auth-type {simple,cognito,delegated,bundle}:

# Server-side Google login
kdcube init --tenant acme --project staging \
  --non-interactive \
  --auth-type bundle \
  --client-id "<google-web-oauth-client-id>" \
  --bootstrap-admin-email admin@example.com

# Cognito
kdcube init --tenant acme --project staging \
  --non-interactive \
  --auth-type cognito \
  --cognito-region eu-west-1 \
  --cognito-user-pool-id <pool-id> \
  --cognito-app-client-id <app-client-id>

To change the method on an already-initialized runtime, use kdcube config apply with the same flags; --dry-run shows the descriptor diff, --restart applies it:

kdcube config apply --tenant acme --project staging \
  --auth-type cognito --cognito-region eu-west-1 \
  --cognito-user-pool-id <pool-id> --cognito-app-client-id <app-client-id> \
  --dry-run

See Authentication modes for per-mode fields.

Typical day-to-day flow

Pass --tenant / --project (or set them with kdcube defaults) to point each command at the runtime you want. --quiet suppresses the banner; the CLI auto-suppresses when stdout is not a TTY and when --json is requested.

# Start the stack
kdcube start --tenant acme --project staging

# Pick up platform code changes (rebuild images + restart)
kdcube refresh --tenant acme --project staging --build
kdcube refresh --tenant acme --project staging --release 2026.5.22.001 --build

# After editing a bundle's config or code — reload without a full restart
kdcube bundle reload <bundle_id> --tenant acme --project staging

# Stop the stack
kdcube stop --tenant acme --project staging

If you've set kdcube defaults --default-tenant acme --default-project staging, you can drop --tenant/--project from these commands entirely.

Runtime flow map

Use these three flows as the mental model for local KDCube operation.

Init is first-time setup for a runtime workdir:

seed descriptors                         platform source
assembly.yaml / bundles.yaml / ...       --path / --upstream / --latest / --release
              |                                |
              +---------------+----------------+
                              v
              kdcube init --descriptors-location <dir> --build
                              |
                              v
       ~/.kdcube/kdcube-runtime/<tenant>__<project>/
       config/*.yaml + repo/ + compose/env files + data/
                              |
                              v
                         kdcube start
                              |
                              v
                  http://localhost:<port>/platform/chat

Refresh is for an already initialized runtime. It preserves staged descriptors:

existing runtime workdir
config/*.yaml  ----------------------------- preserved
      |
      | optional platform source selector
      | none / --path / --upstream / --latest / --release
      v
kdcube refresh [selector] --build
      |
      +-- with --path: copy that local checkout into workdir/repo first
      +-- with --upstream/--latest/--release: update workdir/repo to that ref
      +-- with no selector: rebuild the already staged/recorded source
      +-- with --build: rebuild images
      +-- unless --no-restart: restart the stack

Bundle descriptor apply and reload are bundle-only operations. They do not rebuild platform images or restart Docker:

seed content descriptors
bundles.yaml + bundles.secrets.yaml
              |
              v
kdcube bundle config apply --descriptors-location <dir> [--dry-run]
              |
              v
active runtime bundle descriptors
workdir/config/bundles.yaml + bundles.secrets.yaml
              |
              v
kdcube bundle reload <bundle_id>
              |
              v
proc clears bundle cache and reloads code/config on the next request

Use kdcube config export before replacing live runtime descriptors with an older seed copy. By default it writes bundles.yaml and bundles.secrets.yaml; add --include-platform-descriptors to also export assembly.yaml, secrets.yaml, and gateway.yaml. Local non-git bundle paths are normalized back to host paths, while git-backed entries keep repo/ref/subdir and drop materialized runtime paths. Local host-managed Postgres/Redis entries are exported as descriptor-facing localhost, even though the running containers use host.docker.internal. CLI-managed storage and runtime host paths are exported as null; host_bundles_path is preserved as the source root for unmanaged local bundles. Generated service log_dir: /logs entries are omitted, so a later init can derive them from that runtime's workdir.

kdcube config export \
  --tenant acme \
  --project staging \
  --out-dir /tmp/kdcube-export \
  --include-platform-descriptors

# after review/editing the exported files
kdcube config import \
  --tenant acme \
  --project staging \
  --descriptors-location /tmp/kdcube-export \
  --include-platform-descriptors

config import --include-platform-descriptors treats the reviewed descriptor directory as the local runtime authority and overwrites assembly.yaml, secrets.yaml, and gateway.yaml exactly. It also stages bundle descriptors and regenerates runtime env/config files from the imported platform descriptors. Restart the stack after importing platform descriptors so running services pick up service-level env changes.

Advanced workdir placement

When the runtime must live outside the default base (~/.kdcube/kdcube-runtime), two advanced flags are available on init and refresh:

• --workdir <full-path> — explicit fully-qualified namespaced runtime (trailing path segment must contain __, e.g. /opt/kdcube/acme__staging).
• --workdir-base <base> --tenant T --project P — the CLI composes <base>/<tenant>__<project>/ for you.

# Explicit full path:
kdcube init    --workdir /opt/kdcube/acme__staging
kdcube refresh --workdir /opt/kdcube/acme__staging --build

# Non-default base:
kdcube init --workdir-base /opt/kdcube --tenant acme --project staging

--workdir and --workdir-base are mutually exclusive. Other subcommands (start, stop, reload, bundle, config, info, export) accept the same --tenant/--project shape and a --workdir for the explicit form.

kdcube refresh

Use kdcube refresh to apply platform-side changes (new images, updated SDK code) on an existing initialized workdir. It stops the stack, rebuilds images when --build is given, and restarts — without touching staged descriptors (bundles.yaml, bundles.secrets.yaml, assembly.yaml, secrets.yaml, gateway.yaml):

kdcube refresh --workdir ~/.kdcube/kdcube-runtime/<tenant>__<project> --build

In custom-ui-managed-infra mode, --build also rebuilds proxylogin because delegated-auth behavior is part of the local platform surface.

Refresh also accepts the same platform source selectors as init: --latest, --upstream, and --release <ref>. When you pass --path /path/to/kdcube-ai-app without one of those selectors, refresh restages that local platform source into <workdir>/repo before rebuilding. When you do pass a selector, refresh checks out that selected ref and then uses the staged <workdir>/repo copy. This keeps all compose build contexts aligned with the same source tree while preserving staged descriptors.

refresh refuses if the workdir is not initialized; pair with kdcube init for the first run, then use refresh for every subsequent re-init.

Source builds maintain the Docker host's disposable storage before and after the build. The CLI removes dangling images, including the old generations left when a KDCube latest tag moves, and bounds BuildKit's retained cache while preserving recent layers for the next build. Every maintained local Compose stack rotates Docker-managed container logs at 20 MB with three files per container. The policy uses Docker CLI and Compose contracts and applies to Docker Engine and Docker Desktop on every host OS.

The automatic maintenance does not remove named images, containers, volumes, or data under the runtime workdir. kdcube clean remains the explicit, broader command for removing all unused KDCube image tags and disposable build cache.

Persistent defaults

Save your most-used workdir so you can omit --workdir from every command:

kdcube defaults \
  --default-workdir ~/.kdcube/kdcube-runtime/<tenant>__<project> \
  --default-tenant <tenant> \
  --default-project <project>

Command groups

Lifecycle

Runtime operations

Configuration

kdcube bundle — manage bundles at runtime

Create, update, or delete a staged bundle entry without touching YAML files by hand. Changes are staged and take effect after kdcube bundle reload.

Source mode — point the bundle at a local path or a git repo:

# Local host path under paths.host_bundles_path; CLI stores the /bundles/... path
kdcube bundle <bundle_id> --local-path /Users/you/src/my.bundle

# Already runtime-visible path is also accepted
kdcube bundle <bundle_id> --local-path /bundles/my.bundle

# Git repo (platform clones to /managed-bundles/ on reload)
kdcube bundle <bundle_id> \
  --git-repo git@github.com:org/my-bundle.git \
  --git-ref 2026.4.30

# Git monorepo — bundle lives in a subdirectory
kdcube bundle <bundle_id> \
  --git-repo git@github.com:org/monorepo.git \
  --git-ref main \
  --git-subdir src/my.bundle

Identity and config/secrets patch:

# Set display name, entry module, singleton flag
kdcube bundle <bundle_id> \
  --name "My Bundle" --module entrypoint --singleton

# Patch config and secrets by dotted key path
kdcube bundle <bundle_id> \
  --set-config routines.heartbeat.cron "*/5 * * * *" \
  --set-secret api.token "sk-..." \
  --del-config features.legacy_mode

# Apply all staged changes
kdcube bundle reload <bundle_id>

Normal reload output is concise and operator-facing. Use --verbose only when you need the raw Docker Compose command and full proc response. Use --json for scriptable output.

# Managed deprovision and targeted retirement by bundle ID
kdcube bundle delete <bundle_id>

# Also permit the app hook to delete its durable app-owned data
kdcube bundle delete <bundle_id> --purge-data

# Retire after a failed deprovision attempt and report cleanup incomplete
kdcube bundle delete <bundle_id> --force-retire

Deletion closes the target app in the local proc, stops its scheduled and Data Bus intake, waits for tracked chat tasks, invokes its optional on_app_deprovision(...) hook, removes descriptor authority after success, and retires only that bundle. Other bundles remain loaded. Normal deletion retains durable app data; --purge-data authorizes the hook to remove data it owns. --force-retire is the explicit escape hatch after failed cleanup and reports that cleanup remains incomplete. It does not authorize data purge and does not claim cleanup succeeded. With both flags, a failed purge remains incomplete even though retirement continues. --purge-data requires a running chat-proc, including when combined with --force-retire. See the guarded deprovision contract. kdcube bundle <bundle_id> --delete is the compatibility form of the same complete operation.

Run managed deletion while the target remains declared. Descriptor apply can retire an ID removed from desired inventory, but it does not invoke the app's deprovision hook or authorize durable-data purge. Update reusable seed descriptors after managed deletion succeeds.

Descriptor apply — when a user intentionally edits seed bundles.yaml / bundles.secrets.yaml and wants to reapply that descriptor source of truth to an existing runtime:

kdcube bundle config apply \
  --tenant acme \
  --project staging \
  --descriptors-location /path/to/descriptors \
  --dry-run

kdcube bundle config apply \
  --tenant acme \
  --project staging \
  --descriptors-location /path/to/descriptors \
  --reload

This is not a platform refresh: it touches only bundles.yaml and optional bundles.secrets.yaml in the active runtime config directory. Host local bundle paths from seed descriptors are translated to runtime /bundles/... paths before staging. With --reload, changed declared bundle IDs are reloaded and removed IDs are retired. Removed IDs follow inventory retirement; they do not run on_app_deprovision(...). Each operation targets only its changed ID; unchanged bundles keep running.

Apps own their delegated catalog additions under config.delegated_catalog. After loading or reloading an app, run kdcube bundle catalog check --workdir <runtime-workdir>. The command compares complete descriptor authority with the catalog serving requests and changes nothing. The complete contract is in Current KDCube CLI.

Status — inspect one explicit bundle entry:

kdcube bundle status <bundle_id> --workdir ~/.kdcube/kdcube-runtime/<tenant>__<project>

By default this reports staged descriptor/path/runtime-service diagnostics only and does not list other bundles. For local operator diagnostics, add --live to ask localhost chat-proc to validate that same explicit bundle id:

kdcube bundle status <bundle_id> --live --json --workdir ~/.kdcube/kdcube-runtime/<tenant>__<project>

--live is an operator-level check for someone with local workdir and Docker access. It does not emulate an end-user session or frontend visibility rules.

For scriptable runtime inspection, kdcube info --json emits defaults, the running deployment lock, and runtime mount details when a workdir is selected.

When --local-path or --git-repo is given and the bundle doesn't exist yet, the command creates a new entry (upsert). All other flags require an existing entry. All non-delete flags can be combined in one invocation (single atomic write). --git-ref is required with --git-repo. --git-subdir requires --git-repo.

Full documentation

See additional-README.md in this package or the platform docs:
https://github.com/kdcube/kdcube/blob/main/app/ai-app/docs/service/cicd/cli-README.md