Creator Intelligence
Open-source MCP server · MIT · v0.1.0 release candidate

YouTube research and production planning, served to Claude from your own machine.

Creator Intelligence gives Claude and compatible MCP clients official YouTube evidence, owner-authorized analytics and original production blueprints. It runs locally with TypeScript, Docker Compose, PostgreSQL and Redis, and its workflow ends in a local blueprint JSON export.

Install the MCP View the source on GitHub

Developed by Steven Patrick Buglione as an individual. This website is published from the implementation repository’s docs/ directory.

Developed with YouTube Official API records retain their sources. Creative output is independent.
creator-intelligencetools/list · 18

Evidence

  • creator_status
  • creator_channels
  • creator_videos
  • creator_analytics

Creative workspace

  • creator_ideas
  • creator_blueprints
  • creator_blueprint
  • creator_blueprint_performance

Review records

  • creator_experiments
  • creator_blueprint_export
  • creator_export

Draft writes · one-use owner authorization

  • creator_idea_save
  • creator_blueprint_save

Governed · disabled

  • creator_outliers
  • creator_compare_competitors
  • creator_observed_vph
  • creator_derived_scores
  • creator_extended_public_history
ReadsAuthorized writesReturns DERIVED_METRIC_APPROVAL_REQUIRED
What it does today

From evidence to a production brief

Every record keeps its source, collection time and coverage. Unavailable metrics stay unavailable; nothing is filled in. The application is standalone: it has no external production or publishing integration.

  1. 01

    Research

    Look up public channels and videos through the YouTube Data API. Uploads ingestion is bounded to 200 visible uploads per channel per sync, with quota accounting.

    Public records expire after 30 days.

  2. 02

    Owner reports

    Connect your own channel with read-only OAuth (youtube.readonly, yt-analytics.readonly) to store Analytics and Reporting data.

    Reach reports can stay pending; missing values are never a measured zero.

  3. 03

    Draft

    Save original ideas and version blueprints: hook, outline, script, thumbnail brief, sources and risks. An optional local model can suggest angles.

    Claude’s draft writes need authorization for the exact content.

  4. 04

    Export

    Download a saved revision as creator_intelligence.blueprint.v1 JSON with its ID, revision, export time and provenance.

    A stale revision is rejected.

Included in this release

  • One authenticated MCP server over stdio or Streamable HTTP
  • A local review UI for research, ideas, blueprints and settings
  • Versioned privacy and terms acceptance; MCP sharing off by default
  • Local export, disconnect and deletion of your workspace data

Not part of it

  • Uploading, editing, rendering or publishing videos
  • Scraping, quota bypass or invented competitor analytics
  • Derived scores and extended public-statistical history
  • Sending blueprints to another application
Architecture

One MCP, local services behind it

TypeScript services run with Docker Compose. Default ports bind to loopback. Secrets stay outside Git and Google tokens are encrypted at rest.

Claude or another MCP client

Calls tools. Never receives Google credentials.

stdio launcher or HTTP

MCP server

Same owner and policy checks as the REST API.

localhost:4312/mcp

Fastify API and review UI

Session, CSRF and PKCE controls. Next.js interface for the owner.

API :4311 · UI :4310

Worker

Syncs every six hours, purges hourly, backs up original drafts daily.

BullMQ on Redis :6438

PostgreSQL

Public records and owner reports expire within 30 days; drafts persist until deleted.

:6437

Google API calls come from the API and worker, using your own project’s key and read-only OAuth client. The MCP server reads stored observations and cannot initiate Google collection. No mandatory paid cloud dependency. Optional AI clients may process requested data under their own terms and settings. The project’s MIT license covers its code; dependencies and container images have their own licenses, described in third-party notices.

Install

Run the actual implementation

The repository includes the MCP server, API, worker, review UI, database layer, local blueprint export and tests. The maintained setup helpers target Windows with PowerShell 7.

  1. Prepare your own Google project

    Install the toolchain, then create a dedicated Google Cloud project with YouTube Data API v3, YouTube Analytics API and YouTube Reporting API enabled.

    • Node.js 24.19.x
    • pnpm 11.25.0
    • PowerShell 7
    • Docker Compose

  2. Clone, initialize and start

    git clone https://github.com/StevenBuglione/creator-intelligence.git
cd creator-intelligence
pnpm install --frozen-lockfile
./scripts/init-local.ps1 -OwnedChannelId 'YOUR_CHANNEL_ID' -GoogleProjectNumber 'YOUR_PROJECT_NUMBER'
./scripts/with-local-env.ps1 docker compose up --build -d
./scripts/copy-owner-password.ps1

    Replace the identity placeholders with your values. Follow the setup guide for your API key, read-only OAuth client, channel consent and private credential storage.

  3. Sign in and choose what to share

    Open http://localhost:4310 and sign in with the generated owner password. Review and accept the privacy policy and terms. MCP sharing is a separate choice, off until you enable it. Local drafting works before Google is connected.

  4. Register one connection with Claude Code

    pnpm build
$launcher = (Resolve-Path ./scripts/mcp-launch.ps1).Path
claude mcp add --scope user creator-intelligence -- pwsh -NoProfile -File $launcher

    The local services must be running. The launcher loads your MCP token from private storage. A managed client may need administrator allowlisting. Authenticated Streamable HTTP is also available at http://localhost:4312/mcp; choose one transport.

  5. Verify locally

    ./scripts/with-local-env.ps1 pnpm verify

    Runs lint, TypeScript, unit and integration tests, the production build and desktop/mobile browser workflows against separate test and demo databases with synthetic providers. It makes no Google API calls.

Bring your own Google project, credentials and channel authorization. The source license does not provide access to the developer’s deployment, quota or Google approval.

MCP tools

What the MCP exposes

Eighteen registered tools share the API’s authorization and policy checks. Five governed tools are present only to return an explicit blocked response.

PurposeToolsBehavior
Evidencecreator_statuscreator_channelscreator_videoscreator_analyticsRead
Local state and available official records.
Creative workspacecreator_ideascreator_blueprintscreator_blueprintcreator_blueprint_performanceRead
Original plans and linked owner performance.
Review recordscreator_experimentscreator_blueprint_exportcreator_exportRead
The owner’s workspace records.
Draft writescreator_idea_savecreator_blueprint_saveAuthorized write
Require one-use authorization for the exact draft.
Governed analyticscreator_outlierscreator_compare_competitorscreator_observed_vphcreator_derived_scorescreator_extended_public_historyDisabled
Additional metrics are not implemented or approved.

All calls use the outer argument object {"input": {...}}, for example creator_analytics({"input":{"kind":"overview"}}). Read the tool reference and argument examples. Source text is treated as untrusted data. An MCP request cannot authorize its own draft write. Blueprint export returns the requested revision without sending it to another service.

Release status

A release candidate with visible limits

The repository includes the feature ledger, acceptance plan, automated tests and actual verification results. The 151-feature backlog is a plan, not a list of completed capabilities.

  • Installation-specific setupOwner OAuth and Claude registration require installation-specific setup and verification. Source distribution does not share the maintainer’s credentials.
  • Governed metrics stay offCustom scores and extended public-statistical retention remain disabled pending the relevant accepted YouTube authorization and implementation.
  • Official data onlyNo scraping, quota bypass, invented competitor analytics or automatic YouTube publication.
  • Consent is enforcedThe app enforces versioned privacy/terms acceptance and optional MCP sharing. Google audit review requires deployment-specific evidence and an appropriate reviewer-access arrangement.

See all current limitationsContribute