Bai CLI

./bai CLI Usage Guide

Guide for testing and verifying API endpoints and managing resources with the Backend.AI v2 CLI (./bai).

IMPORTANT:

• ./bai is the v2 REST API CLI. It is separate from the legacy v1 CLI (backend.ai / ./backend.ai).
• Do NOT use v1 CLI commands. Do all testing and verification with ./bai.
• Before running a ./bai command, confirm the command exists in the Entity-Command Reference below. No guessing or fabrication, no searching the CLI source.
• Verify the command tree with --help (works without a server): ./bai {entity} --help, ./bai admin {entity} --help, ./bai my {entity} --help.

v1 → v2 terminology

Entity-Command Reference

Syntax: ./bai [admin|my] {entity} [{sub-entity}] {command} [options]

Each entity is marked by access level — user (user-facing) / admin (superadmin only) / my (own resources). "(empty group)" is a placeholder group with no commands (example: ./bai agent is empty and the actual commands are under ./bai admin agent). Check options with --help.

Core

• domain: user(get) · admin(search, create, update, delete, purge)
• user: user(get, create, update, delete, search) · admin(create, delete, search)
• project: user(get, assign-users, unassign-users · sub role: search) · admin(search, create, update, delete, purge)
• agent: user(empty group) · admin(search, total-resources, update-resource-group)
• image: user(empty group) · admin(search, forget, purge, update · sub alias: create, remove, search)
• entity-label: user(upsert, purge, search)
• session: user(compute-schedule, enqueue, exclude-idle-checks, include-idle-checks, get, logs, project-search, start-service, shutdown-service, terminate, update) · admin(search · sub kernel: search) · my(search)

Compute & Serving

• deployment: user(create, get, update, delete, project-search, chat) · admin(search) · my(search)
  • user sub: access-token(create, get, search, delete, bulk-delete), auto-scaling-rule(create, get, search, update, delete, bulk-delete), replica(search), revision(add, get, current, activate, search), revision-preset(get, search), options(get, replace), chat-cache(show, clear), chat-config(set, show, clear), chat-history(show, clear), policy(empty group)
  • admin sub: policy(search), replica(search), revision(search, refresh), revision-preset(create, get, search, update, delete)
• model-card: user(project-search, get, deploy, available-presets) · admin(search, get, create, update, delete, bulk-delete, scan)
• service-catalog: user(empty group) · admin(search)
• runtime-variant: user(get, search) · admin(get, search, create, update, delete, bulk-delete)
• runtime-variant-preset: user(get, search) · admin(get, search, create, update, delete)
• scheduling-history: sub session / kernel / deployment / route / replica-group — each (search, search-scoped)
• scheduling-handler: admin(list)

Storage

• vfolder: user(my-search, project-search, admin-search, create, project-create, get, upload, download, delete, purge, restore, ls, mkdir, mv, rm, clone, deploy, bulk-delete, bulk-purge)
• vfs-storage: user(create, get, search, list-all, update, delete)
• storage-namespace: user(register, unregister, search, get-by-storage)
• object-storage: user(create, get, search, update, delete)
• storage-host: my(permissions)

Registries & Artifacts

• container-registry: user(empty group) · admin(search, create, update, delete)
• artifact: user(get, update, delete, restore · sub revision: get, approve, reject, cancel-import, cleanup) · admin(search)
• artifact-registry: user(get)
• huggingface-registry: user(create, get, search, update, delete) — no admin variant
• reservoir-registry: user(create, get, search, update, delete)

Access Control & Auth

• entity-type: user(list)
• rbac: sub assignment(assign, revoke, search), entity(search), permission(search), invitation(create, accept, reject, cancel, my-search, my-sent-search, role-search), role(create, get, search, update, delete, project-search, add-permission, remove-permission, replace-permission)
• role: my(search)
• role-preset: admin(create, get, search, update, delete, purge, restore, permission-add, permission-remove, permission-search)
• invitation: admin(search)
• keypair: admin(create, get, search, update, delete · sub ssh: register, get, delete) · my(issue, revoke, search, update, switch-main)
• login-history: admin(search) · my(search)
• login-session: admin(search, revoke) · my(search, revoke)
• login-client-type: user(get) · admin(search, create, update, delete)

Resource Management

• resource-group: user(empty group) · admin(search, get, create, update, delete, resource-info, default-options, default-session-options, allow-domains, allowed-domains, allow-projects, allowed-projects, allow-for-domain, allowed-for-domain, allow-for-project, allowed-for-project)
• resource-allocation: user(project-usage, resource-group-usage) · admin(domain-usage, effective) · my(effective, keypair-usage)
• resource-preset: admin(search, get, create, update, delete, check-availability)
• resource-policy: admin(sub keypair / project / user — each create, get, search, update, delete) · my(keypair, user)
• resource-slot: sub slot-type(search, create, update, delete), agent-resource(search), allocation(search)
• resource-usage: sub domain(search), project(search), user(search)

Monitoring & Audit

• audit-log: user(search)
• fair-share: sub domain / project / user — each (get, search)
• idle-checker-assignment: user(scoped-search, update, purge) · admin(create, search)
• notification: sub channel(get, search, delete), rule(get, search, delete)
• prometheus-query-definition: user(get, search, execute) · admin(create, update, delete, preview)
• prometheus-query-definition-category: user(get, search) · admin(create, delete)
• app-config: my(get) · public(get) — the merged view; public is an app-config-only group for the pre-login read, needing no credentials
• app-config-definition: admin(create, get, search, purge)
• app-config-allow-list: admin(create, get, search, update, purge)
• app-config-fragment: user(get, update, purge, bulk-purge) · my(get, update) · admin(search)
• idle-checker: admin(create, search, update, purge)
• export: admin(list-reports, get-report, audit-logs, keypairs, projects, sessions, sessions-by-project, users, users-by-domain) · my(keypairs, sessions)

Utilities (not entities)

login, logout, config, gql — single commands at the root.

When adding a new CLI command, update this Reference as well (see "Adding a new entity" in client/cli/v2/AGENTS.md).

Setup (Webserver session — recommended)

./bai config set endpoint http://127.0.0.1:8090
./bai config set endpoint-type session
./bai login
# User ID: admin@lablup.com
# Password: (admin password)
./bai config show

Non-interactive environments (CI, Claude Code):

BACKEND_USER=admin@lablup.com BACKEND_PASSWORD=wJalrXUt ./bai login

The session cookie is stored in ~/.backend.ai/session/cookie.dat.

Direct API (alternative)

Access the manager directly without the webserver (HMAC signature auth):

./bai config set endpoint http://127.0.0.1:8091
./bai config set endpoint-type api
./bai config set access-key AKIAIOSFODNN7EXAMPLE
./bai config set secret-key wJalrXUtnFEMI/K7MDENG/bPxRfiCYEXAMPLEKEY
./bai config show

The configuration is stored in ~/.backend.ai/config.toml and credentials.toml.

Command pattern

./bai [admin|my] {entity} [{sub-entity}] {command} [options]

• admin — superadmin-only operations
• my — self-service (the current user's own resources)
• Entity names are singular (domain, user, agent)
• A sub-entity is a Click sub-group (revision, channel, role, etc.)

The standard 6 operations: create, get, search, update, delete, purge (only some, depending on the entity). Special operations: enqueue/terminate (session), revision add/revision activate (deployment), login/logout.

CLI input style

• Default: an individual --option flag per field.
• Secondary: a JSON string or an @file path for complex nested structures (example: --initial-revision, --config).
• Do NOT use raw JSON as a positional argument for create/update (with some admin command exceptions: admin domain create).
• For get/delete/purge, use the entity identifier (UUID, name) as a positional argument.

Search patterns

# admin search — whole system (superadmin only)
./bai admin {entity} search --limit 5

# project-scoped search — the scope ID is a positional argument (not an option)
./bai {entity} project-search {project_id} --limit 5

# self-service search — own resources
./bai my {entity} search --limit 5

Options and filters differ per entity, so check with ./bai {entity} {command} --help.

--order-by syntax

Multi-sort with field:direction:

./bai admin user search --order-by created_at:desc --order-by username:asc

Naming conventions

• The CLI --order-by maps to the DTO order field (common across all entities).
• The CLI --kebab-case option maps to the DTO snake_case field (Click standard).
• Scoped search: the scope ID is a positional argument, not a --scope-* option.

Raw GraphQL

./bai gql (not ./bai admin gql) sends a raw GraphQL query. Useful for testing GQL schema changes or when there is no REST CLI:

./bai gql '{ domain(name: "default") { name } }'
./bai gql --v2 '{ myKeypairs(first: 5) { count edges { node { accessKey } } } }'
./bai gql -f query.graphql
./bai gql --var limit=5 '{ keypair_list(limit: $limit) { items { access_key } } }'

• --v2: targets the Strawberry v2 schema (only needed in direct API mode; session mode provides both)
• --var key=value: query variables (repeatable)
• -f file: read the query from a file
• stdin: echo '{ ... }' | ./bai gql

Common commands

# admin search + filter (superadmin only)
./bai admin domain search --limit 5 --name-contains default
./bai admin user search --status active --order-by created_at:desc
./bai admin agent search --limit 10
./bai admin image search --name-contains python
./bai admin session search --limit 5

# single entity lookup
./bai domain get default
./bai user get <uuid>

# sub-entity operations
./bai admin deployment revision search
./bai rbac role search
./bai notification channel search
./bai scheduling-history session search --limit 10

Testing workflow

Prerequisites

./bai config show   # check endpoint-type
./bai login         # log in if the session expired

After modifying server code

For local development, restart the service first — see the /local-dev skill.

# 1. confirm basic connectivity
./bai admin domain search --limit 1
./bai domain get default

# 2. test the modified entity (with the command matching its level)
#    if user-facing ./bai {entity} ..., if admin-only ./bai admin {entity} ...
./bai admin {entity} search --limit 1
./bai {entity} get {id}

# 3. test the permission boundary
./bai admin {entity} search    # admin succeeds
# after switching to a regular user, the same command → should fail with 403

After commands, verify the runtime behavior with the Grafana MCP — see /observability. To catch errors not surfaced in the CLI response, check Loki ({service_name="manager"} |= "error"); to check the request count, look at Prometheus (backendai_api_request_count).

Testing as a regular user

The default accounts are in fixtures/manager/example-users.json.

# log in as a regular user session
BACKEND_USER=user@lablup.com BACKEND_PASSWORD=C8qnIo29 ./bai login

# should succeed (user-facing)
./bai domain get default

# should fail with 403 (admin only)
./bai admin domain search --limit 1

After testing, revert to the admin credentials.

Smoke test script

Run everything with admin credentials. OK if each command returns JSON.

for cmd in \
  "admin domain search --limit 1" \
  "domain get default" \
  "admin user search --limit 1" \
  "admin project search --limit 1" \
  "admin agent search --limit 1" \
  "admin image search --limit 1" \
  "admin session search --limit 1" \
  "admin resource-group search --limit 1" \
  "audit-log search --limit 1" \
  "rbac role search --limit 1"; do
  echo -n "$cmd: "
  ./bai $cmd 2>&1 | python3 -c "import sys,json;json.load(sys.stdin);print('OK')" 2>&1 || echo "FAIL"
done

When a CLI command is unavailable

If the CLI command for an entity/operation is not in the Reference above:

1. That command is not implemented.
2. Report to the user: "{entity} {operation} is not provided via the CLI. It needs to be implemented."
3. As a temporary workaround, try ./bai gql for operations available via GraphQL.
4. Do NOT guess CLI options or fabricate commands — it wastes time and causes errors.

Related skills

• /local-dev — restart local services before CLI testing
• /observability — check logs/metrics with the Grafana MCP after CLI testing
• /cli-sdk-guide — implement new CLI commands