starlight-creator-mcp

One MCP engine for agentic creators, built so the user's own machine and keys do all the work. Brand packs (Arcanea, GenCreator, Author OS, Anime Legends) sit on top as data; this repository is the unbranded substrate they share.

• BYOK, client-side. Credentials stay on the user's machine in the OS keychain or encrypted-file fallback. Requests go directly to the configured providers.

• Local ledger with provenance. Every generated or imported file is content-addressed (sha256) and carries a .provenance.json sidecar: model, prompt, seed, cost, pack, rubric scores.

• Taste loop. best_of / min_score generate N candidates, score each against the loaded pack's rubric with a vision judge, keep the best.

• Confirmed publishing. Posts go through a signed manifest: prepare → host confirmation → HMAC-signed → publish through the user's Postiz or a local outbox. The server trusts its MCP client's confirmation channel; it does not independently prove human presence. Numerical claims without evidence assets are blocked.

• Packs are data. Canon, rubrics, templates, prompts, styles and routing preferences come from pack.json directories; two packs compose (e.g. author-os + arcanea).

Install

As of 10 October 2026, published @starlight-intelligence/creator-mcp@0.1.0 retains workspace dependencies and fails clean npm dependency resolution. The 0.1.1 repair is a release candidate; it has not been published by this work. Use a reviewed source checkout until the repaired registry release is verified:

git clone https://github.com/frankxai/starlight-creator-mcp.git
cd starlight-creator-mcp
pnpm install --frozen-lockfile
pnpm run build
node packages/bundle/dist/main.js doctor
node packages/bundle/dist/main.js setup --print   # prints Claude / Cursor mcpServers JSON

setup without --print is interactive (stores keys in the OS keychain). Non-TTY correctly exits 2.

Point MCP clients at node /absolute/path/to/packages/bundle/dist/main.js. After the repaired version is published and checked, its npm command is npx --yes @starlight-intelligence/creator-mcp@0.1.1. See release verification and remaining gates.

Providers in this tree: OpenRouter (text, images, async video), MuAPI (226-model image/video catalog), ElevenLabs (speech, sound effects, music), Google Gemini (vision judge). Each activates only when its key is present; doctor runs with none of them. These adapters do not implement transcription.

Storage: local (default, always on), S3-compatible (AWS S3, Cloudflare R2, MinIO — setup --storage), rclone (any remote it supports). S3 endpoints, including AWS, must be listed explicitly in policy.allowHosts; a configured bucket fails closed if the optional connector was omitted (--omit=optional) or its credentials are missing. Publishing media needs a connector that can produce a public URL, so local-only installs can generate and score but not attach images to posts.

Provider output URLs also require an approved host. If a job reports result ingest failed, correct the retrieval/storage error, add only a verified media host to policy.allowHosts if needed, and restart the MCP server to reload the configuration. Call creator_job_status with the same job_id and retry_ingest: true to retrieve the existing provider job again. This submits no new generation. Recovery depends on provider retention and URL validity; polling may carry provider-specific charges. Ordinary status calls leave a failed job unchanged.

Related MCP server: humanMCP

Tools

creator_models · creator_generate_image · creator_generate_video · creator_generate_audio · creator_job_status · creator_score · creator_assets · creator_asset_sync · creator_pack_info · creator_publication_prepare · creator_publication_approve · creator_publish · creator_metrics_ingest · creator_metrics_query · creator_license_status

creator_transcribe is registered only when an injected provider implements the stt capability. It is absent with the default providers.

Every tool has an input schema, an output schema (structuredContent) and annotations. Long jobs return a job_id within wait_ms (≤ 20 s) instead of blocking. Resources: pack://<id>/manifest, pack://<id>/canon/<file>, pack://<id>/templates/<id>, pack://<id>/rubrics/<id>, creator://assets/<sha>, creator://jobs/<id>. Pack prompts become MCP prompts.

Packages

Build a brand on it

import { createCreatorServer } from '@starlight-intelligence/creator-server'
import { createOpenRouterProvider } from '@starlight-intelligence/creator-provider-openrouter'

export const createArcaneaServer = () => createCreatorServer({
  name: 'arcanea-mcp', version: '1.0.0',
  packs: [PACK_DIR],                                   // data only: canon, rubrics, prompts, styles
  providers: vault => [createOpenRouterProvider({ vault })],
  tools: { alias: { creator_generate_image: 'arcanea_generate_image' }, defaults: { creator_generate_image: { style: 'arcanean-cinematic', best_of: 2 } } },
})

Rules: brands depend on the substrate, never on each other; packs are data, not code; one direction of dependency.

Security model

Read SECURITY.md. Short version: keys in the keychain, every outbound fetch goes through a host allowlist that blocks private ranges, every write is sandboxed, publishing requires a human-entered code, and the server never executes shell strings.

Develop

pnpm install
pnpm run build
pnpm test          # unit + protocol-level e2e (spawns the stdio server, mocks providers)
pnpm lint
pnpm inspect       # MCP Inspector against dist/stdio.js

MIT.