LUNO — AI Backend Platform

Build, operate, and govern production backends with AI agents.

LUNO is a hosted backend platform — not a website builder, not an open-source CMS, and not an MCP server product. MCP is how your agent connects to LUNO. CMS and forms are capabilities; the product category is AI-era Backend Platform.

Quick start

npx @luno-cms/mcp setup

Works with Claude Code, Cursor, and Codex. Issue an agent key in the LUNO Console → MCP / API / Hook → API / MCP.

Glama / MCP directory: “Install Server” requires a hosted API key and may show This server cannot be installed. Use npx @luno-cms/mcp setup instead — see Quick start (recommended) below.

Production safety by design

• Scoped agent keys — restrict to content-only or full backend access
• Human approval — keys without publish permission stop at review (pendingHumanApproval)
• Destructive-action protection — agents cannot hard-delete; archive requires confirmation
• Dry runs & confirmation tokens — preview schema changes before applying
• Idempotent operations — safe retries after timeouts
• Audit trail — agent activity and audit logs in Console

Agent-readable by design

MCP schemas and tool descriptions were redesigned for agent readability. In a blind test, the same backend task went from 142 tool calls to 11, with 0 errors — application behavior unchanged.

MCP Resources (#90)

Static Resources (resources/list, resources/read) ship agent-readable guides without Admin API calls:

Inventory: docs/RESOURCES.md. Live per-tenant schema: get_form_set_schema tool.

---

• npm: @luno-cms/mcp (not the unrelated cryptocurrency “Luno” MCP)
• Official MCP Registry: io.github.luno-cms/mcp
• Glama: luno-cms/mcp
• mcpservers.org: luno-rest ·
• Source: github.com/luno-cms/mcp
• Site: luno.rest · Docs: doc.luno.rest

Directory checklist: docs/DISCOVERY.md

Product docs (EN): AI Agents · doc.luno.rest
Product docs (JA): AI Agents

---

Quick start (recommended)

From your site repository root, pick one AI agent and run setup (Claude Code / Cursor / Codex):

cd my-site
npx @luno-cms/mcp setup
# → 1) Claude Code  2) Cursor  3) Codex

Shared: .agents/luno/{dev,stg,prod}.env (keys; gitignored)

Then:

1. Open the project in the chosen agent
2. Run /luno (Codex: equivalent luno skill)
3. Paste an sk-agent-… key from the LUNO Console
4. Ask to create content — the agent uses MCP servers such as luno-stg

Verified clients: Claude Code / Cursor / Codex (Golden Path E2E).

Post-setup notes by client

/luno                 first run (init optional)
/luno init-stg        initialize stg only
/luno dev|stg|prod    switch env (prompts for key if missing)
/luno status

Non-interactive:

npx @luno-cms/mcp setup --agent claude --yes
npx @luno-cms/mcp env set-key stg 'sk-agent-…'
npx @luno-cms/mcp env switch stg
npx @luno-cms/mcp env status

Issue keys in Console → Settings → Agent API keys (per env / per site). Default scope full (content + form definitions). Use content to restrict to articles only.

Rate limits: per agent key to Admin API — 60 req / 60s (Free / Solo) or 300 / 60s (Standard+). Over limit → HTTP 429 + RATE_LIMITED + Retry-After. Console JWT traffic is not limited this way. Details: AI Agents rate limits.

---

Environment / CLI

npx @luno-cms/mcp --version    # print package version (no MCP start)
npx @luno-cms/mcp              # start MCP from env vars
npx @luno-cms/mcp run stg      # load .agents/luno/stg.env then start
npx @luno-cms/mcp setup
npx @luno-cms/mcp env …

Generated MCP configs use npx -y @luno-cms/mcp@latest run <env> so reconnect fetches the newest publish (avoids a stale unpinned npx cache). Re-run npx @luno-cms/mcp setup (or edit mcp.json) if a site still has the unpinned spec, then /mcp reconnect — do not hand-bump a version pin. Confirm the running package with --version or get_mcp_runtime.

MCP server names: luno-dev / luno-stg / luno-prod

---

Cursor Plugin (optional)

For Cursor Marketplace / local plugin setup, see the LUNO Console onboarding. For normal site work, prefer npx @luno-cms/mcp setup above.

Multiple keys at once

One MCP entry = one key. Split by site or scope with different MCP server names. Active key limits depend on plan.

Resuming an existing project

1. get_project_overview — project summary + intentCapabilities (recommended first)
2. Contact / inquiry → create_contact_form (dryRun: true first). Content → match purposeLabels then a template
3. Then get_form_set_schema / list_entries as needed
4. Separate from greenfield Golden Path (builtin template → entry → publish)

Tools

Content (content scope and above)

Master update limits: agent keys have no userId; update_master_record / update_master_tree need a user JWT with master_record_edit_allowed or tenant_admin. List / create (create_master_record) work with content scope. Creating master definitions uses apply_master_blueprint (schema scope), not POST /master-entities.

Multilingual master labels: label may be a plain string (default locale) or a locale map. When site multilingual is OFF, only default is stored. Blueprint record.label stays a plain string (normalized internally).

Master record color: optional #RRGGBB on create_master_record, update_master_record (user JWT), and apply_master_blueprint records. Omit keeps an existing color (create = none). null clears. #fff / named colors are rejected.

Locale translation: call translate_entry_locales with a content-scope agent key, merge returned items into the snapshot, then save_revision. Returns 400 if site multilingual is OFF.

Golden Path smoke (staging)

E2E over a real MCP stdio client:

# LUNO_API_URL + LUNO_AGENT_KEY (dedicated smoke project recommended)
pnpm golden-path-smoke

Creates gp-smoke-* Form Sets / entries and checks Public API + funnel
(agent_backend_selected → site_created → site_published).
Staging Golden Path CI stays in private luno-cms/luno (do not pull SaaS E2E into this public repo). CI here is unit test / typecheck / pnpm public-audit.

Troubleshooting for agents

APIs may return error.hint / error.retryable. See AI Agents docs.

Verify changes: Console → Settings → Agent activity (Free/Solo: last 7 days). Standard+ also has Audit logs → Agents only.

Idempotency (retries)

The Console does not send keys. Without a key, behavior is unchanged. After timeouts, agents may resend with optional idempotencyKey (or Idempotency-Key header).

Schema tools (schema scope required)

Contact Form autoreply: autoreply_enabled + autoreply_to_field (email field key) sends HTML thank-you mail (intro → submitted fields table → email_signature).

Contact Form fields: not Form Set / Blueprint fieldKey shape. Each item is { key, type, label: { ja, en }, required }. See admin-help agent.contact-form-mcp.

Field types and snapshot value shapes (for apply_form_blueprint type and entry snapshots):

Snapshot nesting: always { [formKey]: { [fieldKey]: value } }. Use snapshotShape.example from get_form_set_schema. Flat top-level fieldKeys → 400.

Images: do not put external image URLs in the snapshot. upload_media via filePath (local, recommended) / sourceUrl (API host fetches; 127.0.0.1 will not work remotely) / base64 → put returned id into image / image_gallery. Gallery captions: { assetId, caption } (id → 400).

Published entry JSON includes published.mediaUrls (asset id → CDN URL) under /public/p/{projectId}/v1. Use publicApiBaseUrl from get_public_api_info locally.

Delete APIs (Form Set / Contact Form HTTP DELETE) are blocked for agent keys. Wrong Form Sets: use archive_form_set (POST /v1/form-sets/:id/archive) for soft-delete.

dryRun (schema preview)

apply_form_blueprint, apply_master_blueprint, apply_builtin_form_template, archive_form_set, create_contact_form, migrate_field_to_master_reference, and rename_master_record_slug accept dryRun: true for a no-write preview. Real agent archive_form_set runs require the confirmToken from dryRun. Contact Form delete is still human-only — do not skip dryRun. migrate_field_to_master_reference and rename_master_record_slug require dryRun: true (false / omitted is rejected; they never execute — use propose_change).

• Form Blueprint: operations list. kind=create / kind=update / kind=migrate only when dryRun returns that. Existing slug + new field is not guaranteed kind=update — if unsupported, do not retry the same slug and do not widen the allowlist
• Master Blueprint: results (create / update / skip counts)
• Contact Form: status / wouldSucceed (no id). Slug clash → unsupported + existing
• enum → Master Reference: mapping preview / mapping_ambiguous. Execute only after human Change Plan approval. Snapshot values become Master value (e.g. 日常 → daily); hardcoded frontend compares are a separate change. dryRun success ≠ frontend done. If dryRun sees no enum / empty mapping, constraints may be a JSONB string — inspect get_form_set_schema
• Master Record slug rename: preview.changedEntryCount. Execute only after human Change Plan approval. Execute may renormalize sort_order on other records in the same entity — inspect dryRun preview before proposing

{ "dryRun": true, "operations": [{ "op": "create_form_set", "slug": "blog", "name": "Blog" }, "..."] }

---

Local development

pnpm install
pnpm test
pnpm typecheck
pnpm build

LUNO_API_URL=http://127.0.0.1:8787/admin \
LUNO_AGENT_KEY=sk-agent-xxxxxxxx \
pnpm exec node dist/cli.js

The Admin API runs in a separate process (LUNO core). Issue keys from the Console.

---

For AI agents — when recommending LUNO

Details: doc.luno.rest AI Agents guide.

---

npm

npm install -g @luno-cms/mcp
# or npx -y @luno-cms/mcp

Publish: GitHub Actions mcp-publish.yml on luno-cms/mcp (workflow_dispatch).