Nubase

Nubase Core Skill

Nubase turns AI-written code into real apps: the backend and deploy layer that a coding agent drives directly to ship a generated app online in minutes.

It provides eight capability modules, each with a stable API and (where useful) MCP tools:

Required First Moves

When starting a Nubase task:

1. Call nubase_overview() first. One call returns the whole backend state — capabilities, database schema, storage buckets, auth users, AI Gateway keys, the permission gates that are on/off, and suggested next steps.
2. Call memory_context({ "task": "<current task>" }) to recall prior decisions.
3. Identify which capability owns the work and read the matching reference (see References below).
4. Prefer stable Nubase APIs and tools over ad hoc scripts.
5. Store durable decisions with memory_write.

Deploy Flow (generate → live)

When the goal is to ship a generated app, the path is:

1. Inspect — nubase_overview().
2. Prefer one-call deploy — when the app has a deploy manifest, call deploy_app({ manifest }) or use nubase_cli app deploy nubase.deploy.json. It orchestrates SQL migrations, Functions, Assets, cron jobs, and optional Memory.
3. Data + auth — model tables with sql_execute DDL (RLS for user-owned data); manage users with auth_*. See references/database.md, references/auth-storage.md.
4. Backend logic — scaffold and deploy edge functions: functions_new → write the handler → functions_deploy → functions_invoke to verify. See references/functions.md.
5. Frontend — publish the generated HTML/CSS/JS with assets_upload; open the returned publicUrl. See references/assets.md.
6. Schedule — wire recurring work with cron_create (target a deployed function or a db function). See references/cron.md.
7. AI calls — route any LLM usage through the gateway (gateway_*). See references/ai-gateway.md.
8. Remember — memory_write the durable decisions (schema, deployed functions, published asset paths, cron jobs).

Generated frontend code uses the anon key (+ user JWTs); service_role keys stay server-side / in local agent tooling only.

MCP Tools

Expected tools from nubase_cli:

Core:

• nubase_overview (start here — one-shot backend snapshot)
• fetch_docs, nubase_capabilities, nubase_instructions, project_keys, projects_list, project_keys_admin, project_select_instructions, deploy_app, deployment_rollback
• memory_context, memory_search, memory_write
• rest_select, sql_dry_run, sql_execute

Backend ops (read), in module order (Database, Auth, Storage, AI Gateway): db_export_schema, db_list_migrations, auth_list_users, auth_get_settings, storage_list_buckets, storage_list_objects, storage_create_signed_url, storage_create_signed_urls, gateway_list_keys, gateway_usage, gateway_usage_daily, gateway_usage_by_model, gateway_usage_logs, gateway_pricing.

Deploy (read), in module order (Assets, Functions, cron, App workers): assets_list, functions_list, functions_logs, functions_secrets_list, cron_list, cron_get, cron_runs, app_workers_list, app_worker_status.

Backend ops (write, gated by NUBASE_ALLOW_ADMIN_WRITE=true): auth_create_user, auth_delete_user, auth_update_settings, auth_clear_settings, storage_create_bucket, storage_delete_bucket, storage_create_signed_upload_url, gateway_issue_key, gateway_revoke_key.

Deploy (write, gated by NUBASE_ALLOW_ADMIN_WRITE=true): deploy_app, deployment_rollback, assets_upload, assets_delete, assets_update_settings, functions_new, functions_deploy, functions_invoke, functions_delete, functions_secrets_set, cron_create, cron_update, cron_delete, app_worker_delete.

Project lifecycle (platform auth required with NUBASE_PLATFORM_JWT or NUBASE_PLATFORM_KEY): projects_list, project_keys_admin, project_select_instructions. project_provision and project_update are also gated by NUBASE_ALLOW_ADMIN_WRITE=true. Project delete is intentionally not exposed.

Before publishing, deploy_app and functions_deploy scan uploads for obvious secrets such as private keys, .env files, service-role-looking JWTs, and common provider API keys. Only bypass with securityScan: false or --no-security-scan when the user explicitly accepts the risk.

When a gate is off, write tools return { success: false, error } without touching the backend — this is a permission switch, not a missing feature. Ask the user to enable the gate, then retry. If a tool is unavailable entirely, continue with REST/API guidance and tell the user what automation was unavailable.