Admin and policy

Pin providers, lock endpoints, and govern Skills + Plugins from environment variables — built for managed JupyterHub.

NBI is built to be deployable across a team without surprising users. Every user-facing toggle has a corresponding NBI_* environment variable that lets a platform admin decide its value before anyone signs in.

The policy triad

For boolean capabilities, each policy takes one of three values:

Value Meaning
force-on Always enabled. The user’s Settings dialog shows the control locked on.
force-off Always disabled. The user’s Settings dialog shows the control locked off.
user-choice Honor the user’s config.json (default).

For string-valued settings (provider, model, endpoint) the env var sets the value and the UI control becomes read-only.

Policy categories

  • Mode and feature flags. NBI_CLAUDE_MODE_POLICY, NBI_CLAUDE_CONTINUE_CONVERSATION_POLICY, NBI_CLAUDE_CODE_TOOLS_POLICY, NBI_CLAUDE_JUPYTER_UI_TOOLS_POLICY, NBI_CLAUDE_SETTING_SOURCE_USER_POLICY, NBI_CLAUDE_SETTING_SOURCE_PROJECT_POLICY, NBI_EXPLAIN_ERROR_POLICY, NBI_OUTPUT_FOLLOWUP_POLICY, NBI_OUTPUT_TOOLBAR_POLICY, NBI_REFRESH_OPEN_FILES_ON_DISK_CHANGE_POLICY, NBI_TERMINAL_DRAG_DROP_POLICY, NBI_STORE_GITHUB_ACCESS_TOKEN_POLICY.
  • Provider locks. NBI_CHAT_MODEL_PROVIDER, NBI_CHAT_MODEL_ID, NBI_INLINE_COMPLETION_MODEL_PROVIDER, NBI_INLINE_COMPLETION_MODEL_ID, NBI_CLAUDE_CHAT_MODEL, NBI_CLAUDE_INLINE_COMPLETION_MODEL.
  • Anthropic endpoint pinning. ANTHROPIC_API_KEY, ANTHROPIC_BASE_URL (read by the Claude SDK directly, propagated by NBI).
  • Settings tabs. NBI_SKILLS_MANAGEMENT_POLICY (whole Skills tab), NBI_CLAUDE_MCP_MANAGEMENT_POLICY (Claude MCP Servers tab), NBI_CLAUDE_PLUGINS_MANAGEMENT_POLICY (Claude Plugins tab). Each force-off hides the tab and 403s its REST routes.
  • Skills and Plugins management. NBI_SKILLS_MANIFEST (org-managed YAML manifest URL or path), NBI_SKILLS_MANIFEST_INTERVAL, NBI_SKILLS_MANIFEST_TOKEN, NBI_ALLOW_GITHUB_SKILL_IMPORT, NBI_ALLOW_GITHUB_PLUGIN_IMPORT, NBI_SKILL_MAX_ARCHIVE_MB.
  • Coding-agent launchers. disabled_coding_agent_launchers (traitlet, list of claude-code / opencode / pi / github-copilot-cli / codex), with allow_enabling_coding_agent_launchers_with_env + NBI_ENABLED_CODING_AGENT_LAUNCHERS for per-pod re-enable.
  • Upload staging. NBI_UPLOAD_MAX_MB (default 50), NBI_UPLOAD_RETENTION_HOURS (default 24) govern the staging endpoint used by terminal drag-drop and chat-sidebar attachments.
  • Provider gating. NBI_DISABLED_PROVIDERS (comma-separated list).

Fail-loud resolvers

Every NBI_*_POLICY env var is parsed strictly. A typo like force_on (underscore instead of hyphen) raises at server startup — no silent fallback. That keeps misconfigured production deployments from looking superficially correct.

Skills manifest

An organization can curate a set of Claude Skills via a YAML manifest:

skills:
  - url: https://github.com/your-org/data-skills/tree/main/skills/eda
    scope: user
  - url: https://github.com/your-org/data-skills/tree/main/skills/ml-recipes
    scope: user

Set NBI_SKILLS_MANIFEST=https://your-org/skills.yaml (or a filesystem path). NBI’s reconciler installs every Skill at startup and re-syncs every 24h. Managed Skills are read-only in the UI — users can’t edit, rename, or delete them; if they try to remove one, the reconciler restores it on the next pass.

Reference

Install NBI