Storage adapters

A storage adapter is the layer that persists CMS content. CaretCMS ships four and lets you write your own.

Adapter	Use for	Persistence
Filesystem Default	Local dev, single-server prod, git-tracked content	JSON files on disk
Markdown Content collections	Editing existing Astro content collections in place	YAML frontmatter, plus supported prose in .md files
In-memory	Tests, ephemeral previews	Process memory (lost on restart)
Cloudflare KV	Cloudflare Workers, edge	Cloudflare KV namespace
Custom	Postgres, Redis, S3, anything else	You decide

Picking an adapter

Situation	Pick
Local dev	Filesystem or Markdown (auto-detected)
Site built on Astro content collections (src/content/*.md)	Markdown
Static CDN (static delivery)	Filesystem — .caret/data/ in repo or CI checkout
Single-server prod (Node, Fly, Railway, Render, VPS)	Filesystem with persistent volume
Multi-instance Node	Custom (Postgres / Redis)
Vercel / Lambda	Custom (DB-backed) — no writable filesystem
Cloudflare Workers	Cloudflare KV + R2
Tests	In-memory

The adapters

Filesystem

The default. Writes JSON files under .caret/data/ and metadata under .caretcms/.

• Directory.caret/data/

  • Directorypages/

    • home.json
    • about.json
  • Directorysite/

    • global.json
• Directory.caretcms/

  • revisions.json optimistic-locking counters
  • Directoryhistory/

    • Directorypages/

      • home.json per-entry history (last 50)
  • Directorycollections/

    • products.json dynamic collection metadata

No setup needed — caret() with no options uses it. The directories are created on first write. Customize roots:

astro.config.mjs

import caret, { filesystemStorage } from '@caretcms/core';

caret({
  storage: filesystemStorage({
    dataRoot: 'src/content/cms',         // git-track content in source
    metaRoot: 'src/content/cms/.meta',
  }),
})

When to use: local dev always. Production when you have one server (or shared NFS), don’t need horizontal scaling, and like content in git.

When not to use: serverless (Workers, Lambda — they don’t have writable disk), multi-instance deployments without shared storage, or anything edge-deployed.

Concurrency limits

The adapter serializes writes to its revision-counter file via an in-process queue. Two requests to the same Node process won’t lose increments. Two separate processes can still race — if you run multiple instances, use KV or a custom adapter with real locking.

Markdown

For sites already built on Astro content collections. Instead of a separate .caret/data/ store, this adapter reads and writes the Markdown files you already have under src/content/. Field edits update YAML frontmatter. Editors can also update supported prose blocks on the rendered page and publish those changes back to .md files.

• Directorysrc/content/

  • Directoryblog/

    • hello-world.md frontmatter = entry data
    • second-post.mdx
  • Directoryauthors/

    • jane.md
• Directory.caretcms/ CMS state stays out of your content

  • revisions.json
  • Directoryhistory/

    • …
  • Directorycollections/

    • …

One set of files powers both Astro’s native getCollection() / render() and CaretCMS editing — there’s no copy or sync step. Wire it explicitly:

astro.config.mjs

import caret, { markdownStorage } from '@caretcms/core';

caret({
  storage: markdownStorage({
    contentRoot: 'src/content',   // default
  }),
})

Auto-detected — you usually don't set this

When you haven’t specified storage and your project has content collections under src/content/, CaretCMS selects markdownStorage() for you — so the Studio lists your collections immediately instead of showing “No collections yet”. Pass an explicit adapter to override; storage: filesystemStorage() opts back out.

How edits are written: Studio and normal data-caret fields update frontmatter. Supported prose edits are kept in a private draft until Publish replaces the matching source range. Unedited body text stays byte-for-byte unchanged. Revisions, history, and collection metadata live in the .caretcms/ sidecar, so your source files stay free of CMS bookkeeping.

When to use: any site whose content is already Markdown content collections, when you want to edit those entries’ fields — in Studio or inline — without moving them into a separate store. Pairs naturally with registered schemas derived from the same Zod schema your content.config.ts already uses.

Drafts and publish

Authenticated editors edit through a per-editor draft overlay (under .caret/drafts/), so your source .md files stay unchanged until an explicit Publish writes the draft to frontmatter or prose.

See Markdown body editing for supported blocks, formatting, conflict handling, and Astro version notes.

Static delivery note: fields rendered through getCollection() at build time only refresh after rebuild. Pair with commit-on-publish + CI when using static delivery.

v1 boundaries

• Flat collection directories only (no nested-slug subfolders yet).
• .md prose is editable; .mdx, JSX, tables, code blocks, and raw HTML blocks stay read-only.
• No demo/sandbox mode: this adapter intentionally doesn’t implement per-session overlays for anonymous visitors (that would mean writing to your source files). Use the filesystem or KV adapter for demo mode.

In-memory

Holds everything in JS maps. Lost when the process restarts.

When to use: unit and integration tests. Anywhere you don’t want disk writes polluting state across runs.

Unlike the other adapters, in-memory isn’t wired through caret({ storage }) — there’s no provider entrypoint to reference. You construct InMemoryAdapter directly, the way the package’s own test suite does:

import { InMemoryAdapter } from '@caretcms/core';

const adapter = new InMemoryAdapter();
adapter.preload('pages', [{ id: 'home', data: { headline: 'Hello' } }]);

Cloudflare KV

Install the Cloudflare package:

npm install @caretcms/cloudflare

astro.config.mjs

import { defineConfig } from 'astro/config';
import cloudflare from '@astrojs/cloudflare';
import caret from '@caretcms/core';
import { cloudflareStorage, r2Uploads } from '@caretcms/cloudflare';

export default defineConfig({
  output: 'server',
  adapter: cloudflare(),
  integrations: [
    caret({
      storage: cloudflareStorage({ binding: 'CMS_KV' }),
      uploads: r2Uploads({ binding: 'CMS_R2' }),
    }),
  ],
});

Bindings come from wrangler.toml:

wrangler.toml

name = "my-site"
compatibility_date = "2025-01-01"

[[kv_namespaces]]
binding = "CMS_KV"
id = "<your-kv-namespace-id>"

[[r2_buckets]]
binding = "CMS_R2"
bucket_name = "my-site-uploads"

Storage layout in KV:

collections                    → ["pages", "site", "products"]   (known-collections index)
index::pages                   → ["home", "about"]                (per-collection entry index)
pages::home                    → entry JSON
rev::pages::home               → revision counter
history::pages::home           → revision history array
meta::products                 → CollectionMetadata for dynamic collections

Bundled seed content

By default (bundledFallback: true), when a key isn’t in KV the adapter falls back to JSON bundled from .caret/data/**/*.json at build time. Commit starter content there and it ships with the Worker, serving as the initial value until an edit writes to KV. Pass cloudflareStorage({ binding: 'CMS_KV', bundledFallback: false }) to turn it off.

When to use: any Cloudflare Workers deploy, edge-distributed sites, anything that scales horizontally.

When not to use: non-Cloudflare hosts. KV’s eventual consistency (60s replication) is fine for content but not for tightly-coupled state.

See Deployment → Cloudflare for the end-to-end setup.

Custom

Implement the StorageAdapter interface:

src/cms/postgres-adapter.ts

import type {
  StorageAdapter,
  EntryData,
  HistoryEntry,
  CollectionMetadata,
} from '@caretcms/core';

export class PostgresAdapter implements StorageAdapter {
  // collection / entry reads
  async discoverCollections(): Promise<string[]> { /* ... */ }
  async isKnownCollection(collection: string): Promise<boolean> { /* ... */ }
  async getEntry(collection: string, id: string): Promise<EntryData | null> { /* ... */ }
  async listEntryIds(collection: string): Promise<string[]> { /* ... */ }
  async listEntries(collection: string): Promise<EntryData[]> { /* ... */ }

  // entry mutations
  async writeEntry(collection: string, id: string, data: Record<string, unknown>): Promise<void> { /* ... */ }
  async deleteEntry(collection: string, id: string): Promise<void> { /* ... */ }

  // optimistic locking
  async getRevision(collection: string, id: string): Promise<number> { /* ... */ }
  async bumpRevision(collection: string, id: string): Promise<number> { /* ... */ }

  // history
  async getHistory(collection: string, id: string): Promise<HistoryEntry[]> { /* ... */ }
  async appendHistory(collection: string, id: string, entry: HistoryEntry): Promise<void> { /* ... */ }

  // dynamic collection metadata
  async createCollection(metadata: CollectionMetadata): Promise<void> { /* ... */ }
  async deleteCollection(collection: string): Promise<void> { /* ... */ }
  async getCollectionMetadata(collection: string): Promise<CollectionMetadata | null> { /* ... */ }
  async listCollectionMetadata(): Promise<CollectionMetadata[]> { /* ... */ }

  // optional — only for demo/sandbox mode (per-session write isolation)
  // async makeSessionOverlay(sessionId: string): Promise<StorageAdapter> { /* ... */ }
}

Reordering isn’t a separate adapter method — the mutation engine persists order by writing entries with an updated order field and bumping revisions, so you don’t implement a reorderEntries call.

Wire it via a provider reference:

astro.config.mjs

import caret, { defineStorageProvider } from '@caretcms/core';

export default defineConfig({
  integrations: [
    caret({
      storage: defineStorageProvider({
        entrypoint: './src/cms/postgres-adapter.ts',
        exportName: 'postgresStorageProvider',
        options: { connectionString: process.env.DATABASE_URL },
      }),
    }),
  ],
});

The integration imports your entrypoint at runtime, calls the named export, and uses whatever it returns. The export can be:

• A factory function — called with options once, returns an adapter
• An adapter instance — used directly

Concurrency contract

bumpRevision(collection, id) must atomically increment and return the new value. The mutation engine relies on this for optimistic locking.

Implementation	Atomicity strategy
Filesystem	Serialized in-process queue (single process only)
KV	Single-key writes, last-write-wins (acceptable in practice for content)
Postgres	UPDATE ... RETURNING in a single statement
Redis	INCR

Caution

If your backend can’t guarantee atomicity, you’ll see lost updates under contention.

Field validation

Field paths come from user input. The mutation engine already rejects prototype-pollution attempts (__proto__, constructor.prototype, etc.) before they reach your adapter. You don’t need to re-validate; just trust the inputs you receive in adapter methods.

Next steps

DeploymentAdapter choices per host.

API ReferenceWhat each adapter method serves.

Configurationcaret({ storage }) options.