CodedByKay.ArtShow.Mcp

MCP server that does the whole CodedByKay.ArtShow.CLI workflow — add, edit, view, and delete artwork; manage classifiers; publish artwork.json; check catalog/R2 sync — from an MCP client (Claude Code, Claude Desktop, etc) instead of the CLI's interactive TUI. Batch-capable throughout (add/update/ delete many artworks in one call) and can classify new artwork itself when you don't hand it category/medium/tags/groups (see "Classification" below).

This is a second writer of artwork.db alongside CodedByKay.ArtShow.CLI itself, and — new as of this rebuild — a second writer of R2 objects and of the site's data/artwork.json too. See ArtShow.Workspace/working-on/mcp-rebuild-plan.md for the full rationale and ArtShow.Workspace/AGENTS.md for how this fits the rest of the workspace. Both this server and the CLI talk to the same SQLite file in WAL mode, which supports that safely (busy_timeout + a SAVEPOINT per batch item here).

Uses Node's built-in node:sqlite (no native DB dependency to build), plus sharp for imaging and @aws-sdk/client-s3 for R2 — both installed from prebuilt binaries, no native build step. Requires Node.js 22.5+.

Setup

npm install
npm run build

R2 credentials aren't configured separately — this server reads the exact same sources the CLI does (see "Configuration" below), so if dotnet run in CodedByKay.ArtShow.CLI already works, this does too.

Related MCP server: serve-mcp

Tools

Read / inspect

• list_artworks — filter (ids, type, category, medium, tag, group, text search, date range, missing-R2-key) + sort + paginate.

• get_artwork — fetch one or more artworks by id.

• get_status — no network: catalog counts, classifier breakdown, the pending-publish flag, and the resolved config (paths, R2 bucket, and where credentials were found or that they're missing). Good first call.

• check_sync — cross-references the catalog against the R2 bucket: keys referenced but missing, bucket objects orphaned (unreferenced), total bytes. verifyUrls: true also HEAD-checks every public URL.

• verify_catalog — DB-only consistency sweep the CLI has no equivalent of: missing R2 keys/video ids, invalid dates, empty titles, duplicate R2 keys, classifier values used by an artwork but not in the Classifiers table.

Write — artworks

• add_artworks — the centerpiece. Add one or more artworks (images and/or YouTube videos, mixed freely) in a single call. Per image: EXIF- orients, encodes a full-res WebP (q90) + 600w WebP thumbnail (q75), allocates collision-free R2 keys, uploads both, inserts the row, marks the catalog dirty, HEAD-verifies both public URLs, then deletes the local source file (deleteSourceAfterUpload, default true — only deletes after both URLs verify live). Every item succeeds or fails independently. dryRun: true validates/encodes/resolves classifiers without uploading or writing anything. See "Classification" below for onMissingClassifiers and classifierMode.

• update_artworks — batch partial update by id; nullable fields (description/createdDate/category/medium/r2Key/thumbR2Key/ videoId) explicitly settable to null to clear them. Doesn't touch R2 or re-encode images.

• delete_artworks — batch delete, including R2 objects (the CLI's real "Delete artwork" behavior) unless keepFiles: true. Requires confirm: true.

• replace_artwork_image — point an existing image row at a new local file: uploads new R2 objects, updates the row, verifies, then removes the old R2 objects. The one workflow the CLI has no direct answer for short of delete-and-re-add.

• relocate_artwork_media — batch-move an image row's R2 objects to their canonical originals/<year>/<slug>.webp / thumbs/<year>/<slug>-600w.webp path (year from CreatedDate, falling back to AddedDate) via a server-side R2 copy — no download, no re-encoding, bytes untouched. For when CreatedDate was corrected after upload and left the object under the wrong year folder. A row already at its canonical keys is left alone and reported unchanged; a manual copy done ahead of time (new key already present, old key already gone) is tolerated and still updates the row. Every id succeeds or fails independently.

Classifiers

• list_classifiers — the vocabulary (Category/Medium/Tag/Group), each with a usage count and example titles. Pass match: string[] to fuzzy-match candidate values against it (surfaces "watercolour" → "watercolor" instead of letting a near-duplicate spelling become a second classifier).

• create_classifiers — batch add known values without attaching them to any artwork.

• rename_classifier — cascades into every artwork using it. merge: true folds a rename into an already-existing classifier of the same name instead of erroring — the real fix for a near-duplicate.

• delete_classifier — reports usage count; force: true to unlink from artworks and delete anyway.

Publish

• publish — writes data/artwork.json byte-identically to what the CLI's System.Text.Json writer produces (see "Publish parity" below), clears the pending-publish flag, and returns a diff (ids added/removed, and which fields changed on modified ids). Metadata only — does not commit or push; review the diff, then commit + push the site repo yourself (Netlify auto-deploys on push). dryRun: true computes the diff without writing.

• set_publish_state / get_publish_state — read/manually set the pending-publish flag.

Maintenance

• prune_orphaned_r2_objects — deletes bucket objects no catalog row references. Dry-run by default; confirm: true to execute.

Also: an MCP resource artshow://classifiers (the same vocabulary list_classifiers returns, without a tool call) and an MCP prompt add-new-drawings that walks through classifying-and-adding every image in a staging folder end to end.

Classification

This server can't see an image — classification is the calling agent's job. Three things make that consistent instead of ad hoc:

1. list_classifiers is the vocabulary, with usage counts and example titles, so an agent picks an established value over inventing a new one.

2. onMissingClassifiers on add_artworks (default "error"): an item with no category/medium/tags/groups at all is rejected, and the error carries the entire vocabulary — so an agent that hits it can look at the image and retry immediately instead of a second round trip. "warn" inserts anyway with a flag on the result; "allow" is silent (the pre-rebuild behavior).

3. classifierMode: "lenient" (default, matches the CLI's own auto-add-on-first-use behavior) accepts a new value and creates it, reporting it in the result as classifiersCreated. "strict" rejects a value not already in the vocabulary — if something close exists ("watercolour" vs "watercolor") the error suggests it by name; if nothing does, it says so and points at list_classifiers.

Publish parity

CodedByKay.ArtShow.CLI writes data/artwork.json via JsonSerializer.Serialize(items, new JsonSerializerOptions { WriteIndented = true }). For this server's publish to be a safe drop-in replacement (rather than having every alternate publish between the CLI and this server rewrite the whole file), src/publishing/dotnet-json-encoder.ts reimplements that writer's exact output byte-for-byte: CRLF line endings, 2-space indent, no trailing newline, comma-before-newline placement, empty arrays/objects collapsed to []/{}, and JavaScriptEncoder.Default's escaping (&, ', <, >, and everything above 0x7E escaped as uppercase \uXXXX). Verified with a byte-for-byte round-trip test against the live file — see src/test/dotnet-json-encoder.test.ts.

Configuration

Reads the same sources CodedByKay.ArtShow.CLI does (mirroring Program.cs's ConfigurationBuilder precedence — env vars win, then dotnet user-secrets, then appsettings.json), so no separate credential setup is needed if the CLI already works:

Env var overrides (all optional):

Call get_status to see exactly what was resolved and where each value came from (including whether credentials were found and in which file).

Safety

• Every destructive tool (delete_artworks, prune_orphaned_r2_objects) requires confirm: true; both plus add_artworks/replace_artwork_image support dryRun.

• A local source file is never deleted before both its R2 public URLs HEAD 200.

• If a catalog insert fails after an R2 upload already succeeded, the server best-effort deletes the just-uploaded objects rather than leaving them orphaned.

• This server never runs EF Core migrations and never creates tables — a schema mismatch at startup fails loudly and tells you to run the CLI once instead of silently improvising.

• Each batch item (in add_artworks/update_artworks) runs in its own SQLite SAVEPOINT, so one bad item can't corrupt the catalog or block the rest of the batch.

Registering with a client

{
  "mcpServers": {
    "artshow-catalog": {
      "command": "node",
      "args": ["C:\\Users\\kaywi\\dev\\CodedByKay.ArtShow.Mcp\\dist\\index.js"]
    }
  }
}

Or with the Claude Code CLI:

claude mcp add artshow-catalog -- node C:\Users\kaywi\dev\CodedByKay.ArtShow.Mcp\dist\index.js

Development

npm run dev      # tsc --watch
npm test         # build, then run src/test/*.test.ts under node --test

Several tests read the live artwork.db and site artwork.json read-only (never mutate) to check real-world parity — they skip gracefully if those paths don't exist on the machine running them.