System Architecture

Genie Helper is a multi-service platform that combines AnythingLLM, Directus CMS, Ollama, and Stagehand browser automation into a unified AI operations platform for content creators.

Architecture Overview

Browser → geniehelper.com (React SPA)
  → /app/*          authenticated creator dashboard
  → /api/directus/  Directus REST (port 8055) — data layer
  → /api/llm/       AnythingLLM (port 3001)   — chat + agent + embed widget

AnythingLLM Agent
  → directus MCP   (17 tools): CRUD collections, trigger flows, manage users/files
  → ollama MCP     (3 tools):  generate, chat, list-models
  → stagehand MCP  (9 tools):  browser sessions, navigate, act, extract, cookies, screenshot
  → Action Runner: [ACTION:slug:{"params"}] tag interceptor → pre-built flows

Media Worker (BullMQ)
  → scrape_profile  — Stagehand OF login + data extraction
  → publish_post    — Stagehand-based cross-platform posting
  → apply_watermark — ImageMagick watermarking
  → create_teaser   — FFmpeg video preview generation
  → post_scheduler  — polls scheduled_posts every 60s

Service Layout

AnythingLLM

Port: 3001
PM2 Name: anything-llm
Role: Chat API, AI agent, embed widget
Files: server/, storage/

Directus CMS

Port: 8055
PM2 Name: agentx-cms
Role: Collections, auth, REST API
Files: cms/

Stagehand

Port: 3002
PM2 Name: stagehand-server
Role: Browser automation
Files: scripts/stagehand-mcp-server.mjs

Dashboard

Port: 3100
PM2 Name: genie-dashboard
Role: React SPA (serve dashboard/dist/)
Files: dashboard/

Media Worker

Port: —
PM2 Name: media-worker
Role: BullMQ consumer (Redis)
Files: media-worker/

Collector

Port: —
PM2 Name: anything-collector
Role: Document ingestion
Files: AnythingLLM collector scripts

Ollama

Port: 11434
PM2 Name: (system)
Role: Local LLM inference
Files: System service

MCP Servers (29 Tools Total)

Genie Helper uses the Model Context Protocol to give the AI agent access to three core systems:

Directus MCP (17 Tools)

Script: scripts/directus-mcp-server.mjs

View All Directus Tools

list-collections — List all available collections
get-collection-schema — Get schema for a specific collection
read-items — Read multiple items from a collection
read-item — Read a single item by ID
create-item — Create a new item in a collection
update-item — Update an existing item
delete-item — Delete an item
search-items — Search items with filters
trigger-flow — Trigger a Directus Flow
get-me — Get current user info
list-users — List all users
get-user — Get a specific user
update-user — Update user details
create-user — Create a new user
list-files — List uploaded files
get-file — Get file details
list-flows — List all automation flows

Ollama MCP (3 Tools)

Script: scripts/ollama-mcp-server.mjs

generate — Generate text completions
chat — Interactive chat with model
list-models — List available local models

Stagehand MCP (9 Tools)

Script: scripts/stagehand-mcp-server.mjs

View All Stagehand Tools

start-session — Initialize browser session
navigate — Navigate to URL
act — Perform browser action (click, type, etc.)
extract — Extract data from page
observe — Observe page state
close-session — Close browser session
set-cookies — Set browser cookies
get-cookies — Get browser cookies
screenshot — Capture screenshot

MCP Configuration

Config File: storage/plugins/anythingllm_mcp_servers.json
Auto-boot: Patched server/utils/boot/index.js → bootMCPServers() on startup

Data Layer (Directus Collections)

Genie Helper uses Directus as its primary data store with 11+ key collections:

Collection	Purpose	Key Fields
`creator_profiles`	Platform accounts, encrypted credentials, scrape status	`platform`, `username`, `credentials` (encrypted)
`scraped_media`	Content with engagement metrics	`url`, `views`, `likes`, `revenue`
`scheduled_posts`	Post queue (worker polls every 60s)	`scheduled_at`, `platforms`, `status`
`media_jobs`	BullMQ job records	`job_type`, `status`, `result`
`hitl_sessions`	Human-in-the-loop login requests	`platform`, `status`, `completed_at`
`platform_sessions`	Encrypted browser cookies from extension	`platform`, `cookies` (encrypted)
`taxonomy_dimensions`	6 super-concept content classification	`dimension_name`, `weight`
`taxonomy_mapping`	3,208 classified tags across 6 concepts	`tag`, `dimension`, `confidence`
`fan_profiles`	Fan engagement data	`username`, `ltv`, `last_interaction`
`action_flows`	Action Runner flow definitions	`slug`, `steps` (JSON)
`agent_audits`	Every ACTION execution logged	`action_slug`, `success`, `error`

All platform credentials are encrypted with AES-256-GCM using server/utils/credentialsCrypto.js before storage.

Action Runner System

The Action Runner is a middleware layer that intercepts special tags from the AI model’s output:

How It Works

Model Emits Action Tag

The AI model outputs: [ACTION:taxonomy-tag:{"media_id":123}] in its prose

Action Runner Intercepts

The action-runner plugin detects the tag and strips it from visible chat

Flow Execution

Looks up taxonomy-tag in action_flows collection and executes the steps JSON

Status Streaming

Streams execution status back to the chat interface in real-time

Audit Logging

Records execution to agent_audits (success/error/miss)

Available Actions

{
  "slug": "taxonomy-tag",
  "steps": [
    {
      "type": "mcp_call",
      "server": "directus",
      "tool": "read-item",
      "params": {"collection": "scraped_media", "id": "{{media_id}}"}
    },
    {
      "type": "mcp_call",
      "server": "ollama",
      "tool": "generate",
      "params": {"model": "scout-fast-tag", "prompt": "Classify: {{item.title}}"}
    },
    {
      "type": "mcp_call",
      "server": "directus",
      "tool": "update-item",
      "params": {"collection": "scraped_media", "id": "{{media_id}}", "data": {"tags": "{{classification}}"}}
    }
  ]
}

Action Slug	Purpose	MCP Tools Used
`scout-analyze`	Scrape URL + AI analysis	`stagehand:navigate`, `stagehand:extract`, `ollama:generate`
`taxonomy-tag`	Auto-classify content with taxonomy	`directus:read-item`, `ollama:generate`, `directus:update-item`
`post-create`	Draft platform-specific post	`directus:read-items`, `ollama:chat`, `directus:create-item`
`message-generate`	Fan engagement message	`directus:read-item`, `ollama:generate`
`memory-recall`	Search stored data + summarize	`directus:search-items`, `ollama:generate`
`media-process`	Queue media job	`directus:create-item` (media_jobs)

Media Worker (BullMQ)

The Media Worker is a BullMQ consumer that processes asynchronous jobs: Queues:

scrape-jobs — Profile scraping
media-jobs — Media processing
onboarding-jobs — User onboarding workflows

Job Types:

scrape_profile

Uses Stagehand to log into OnlyFans, extract stats and content, store in scraped_media

publish_post

Cross-platform posting via Stagehand browser automation

apply_watermark

ImageMagick-based image watermarking (100ms per operation)

create_teaser

FFmpeg video preview generation (~30s CPU per clip)

post_scheduler

Polls scheduled_posts every 60s and publishes ready posts

Worker File: media-worker/index.js
Redis Backend: Required for BullMQ job queue

Dashboard (React SPA)

Path: dashboard/
Stack: React 18 + Vite + Tailwind CSS + React Router v6
Doc Root: dashboard/dist/ served at geniehelper.com/

Routes

Public Routes:

/ — Home
/pricing — Pricing plans
/about — About page
/register — Registration (invite-gated)
/login — Login

Authenticated Routes (/app/*):

/app/dashboard — Main dashboard
/app/media — Media library
/app/calendar — Post calendar
/app/fans — Fan management
/app/analytics — Performance analytics
/app/platforms — Platform connections
/app/settings — User settings

Admin Routes:

/admin — Directus + AnythingLLM iframes
/view-as — User impersonation

AnythingLLM embed widget injected bottom-right on all /app/* routes:

Embed ID: cf54a9c0-224c-469d-b97b-5dc8095eac82 (Administrator workspace)
Greeting: Set via data-greeting attribute on script tag
Component: dashboard/src/components/AgentWidget/index.jsx
Trigger: window.__genieOpenChat()

API Proxies

Nginx Configuration:

/api/directus/ → http://localhost:8055
/api/llm/ → http://localhost:3001

Authentication:

Directus JWT (auto-refresh)
sessionStorage for impersonation tabs

Registration:

Invite-gated — Alpha invite code validated against AnythingLLM invite API

Data Flow Examples

Profile Scraping Flow

User Clicks 'Scrape Profile'

Dashboard sends POST to /api/queue/scrape-profile with creator_profile_id

Job Enqueued

BullMQ job created in scrape-jobs queue, record added to media_jobs

Media Worker Picks Up Job

Worker checks platform_sessions for encrypted cookies

Stagehand Session Started

If cookies exist: set-cookies + navigate to platform
If cookies missing: Create hitl_sessions record (yellow dashboard banner)

Data Extraction

Stagehand extract tool scrapes stats, content URLs, engagement metrics

Data Storage

Items created in scraped_media collection via Directus API

Job Completion

media_jobs record updated with status and result count

Content Classification Flow

User Asks 'Tag this content'

Chat message sent to AnythingLLM agent

Agent Emits Action

Model outputs: [ACTION:taxonomy-tag:{"media_id":123}]

Action Runner Intercepts

Looks up taxonomy-tag flow in action_flows collection

Step 1: Read Media

MCP call: directus:read-item from scraped_media

Step 2: AI Classification

MCP call: ollama:generate with scout-fast-tag model

Step 3: Update Tags

MCP call: directus:update-item with classified tags

Audit Log

Execution recorded in agent_audits collection

Security Architecture

Credential Encryption

File: server/utils/credentialsCrypto.js

// Encrypted credentials stored as:
{
  "enc": "v1:iv:tag:ciphertext"
}

// Functions:
encryptJSON(plainObject) → { enc: "v1:..." }
decryptJSON(encObject) → plainObject

CREDENTIALS_ENC_KEY_B64=<base64 of 32 bytes>
DIRECTUS_ADMIN_TOKEN=<admin-token>
RBAC_SYNC_WEBHOOK_SECRET=<webhook-secret>

Key Points:

AES-256-GCM encryption
No encryption keys in browser — all crypto is server-side only
Handles both new object format and legacy raw strings

Authentication Flow

User logs in via /login (React form)
Credentials sent to /api/directus/auth/login
Directus returns JWT access token + refresh token
Dashboard stores tokens in localStorage
All API requests include Authorization: Bearer <token>
Auto-refresh when access token expires

RBAC Synchronization

User roles are synced between Directus and AnythingLLM: Endpoint: server/endpoints/api/rbacSync.js
Webhook: Directus fires webhook on user create/update
Secret: RBAC_SYNC_WEBHOOK_SECRET validates webhook authenticity

Nginx Reverse Proxy

Plesk Users: All nginx changes via Plesk UI only. Never add location / — Plesk generates its own, duplicate causes startup failure.

Current proxy layout (full config in docs/nginx/geniehelper.com.vhost_nginx.conf):

# SPA fallback for React Router
error_page 404 =200 /index.html;

location /api/directus/ {
    proxy_pass http://127.0.0.1:8055/;
    proxy_set_header Host $host;
}

location /api/llm/ {
    proxy_pass http://127.0.0.1:3001/;
    proxy_http_version 1.1;
    proxy_set_header Upgrade "upgrade";
    proxy_set_header Connection "upgrade";
    proxy_set_header Host $host;
    proxy_read_timeout 86400;
}

Critical rules:

Sub-path locations (/app/, /api/llm/) and exact-match (= /path) are safe
SPA routing: error_page 404 =200 /index.html;
WebSocket: use "upgrade" literal for Connection header

Project Structure

agentx/
├── scripts/          MCP servers + CLI utilities
├── server/           AnythingLLM backend (Express + Prisma)
│   ├── endpoints/api/
│   │   ├── captions.js          POST /api/captions/generate
│   │   ├── credentials.js       AES-256-GCM credential endpoints
│   │   ├── impersonate.js       Admin impersonation
│   │   ├── queue.js             BullMQ job management
│   │   ├── rbacSync.js          Directus ↔ AnythingLLM user sync
│   │   └── register.js          Invite-gated registration proxy
│   └── utils/
│       ├── actionRunner/        Action flow executor
│       ├── credentialsCrypto.js AES-256-GCM encryption
│       └── boot/                MCP server auto-boot
├── dashboard/        React SPA (Vite + Tailwind + React Router v6)
│   └── src/
│       ├── pages/               Route pages
│       ├── components/Layout/   AppLayout (embed widget), Sidebar, PublicLayout
│       └── utils/api.js         Directus + LLM API client
├── media-worker/     BullMQ media processing worker
├── cms/              Directus extensions
├── storage/          AnythingLLM persistent storage + MCP config
├── docs/nginx/       Nginx vhost configs (Plesk reference)
└── .claude/sessions/ Development session logs

Performance Considerations

Server Specs: $100/mo IONOS dedicated VPS, self-hosted

Resource Usage:

LLM inference: CPU-bound Qwen2.5 7B ~2-5s/call, ~4.8GB RAM pinned
Stagehand sessions: ~300MB RAM/active browser → ~33 concurrent sessions hard ceiling (10GB available)
FFmpeg clip generation: ~30s CPU per clip (real bottleneck)
Sharp/ImageMagick watermark: ~100ms (effectively zero cost → unlimited justified)

Scaling Limits:

Current Plesk VPS is CPU-only
dolphin3:8b and any >7B model will stall on CPU
Production upgrade: GPU VPS or stick with qwen-2.5:latest

Next Steps

Quick Start

Install and configure Genie Helper on your server

API Reference

Explore the Directus and AnythingLLM API endpoints

MCP Servers

Learn how to extend MCP servers with custom tools

Deployment

Production deployment best practices

​System Architecture

​Architecture Overview

​Service Layout

AnythingLLM

Directus CMS

Stagehand

Dashboard

Media Worker

Collector

Ollama

​MCP Servers (29 Tools Total)

​Directus MCP (17 Tools)

​Ollama MCP (3 Tools)

​Stagehand MCP (9 Tools)

​MCP Configuration

​Data Layer (Directus Collections)

​Action Runner System

​How It Works

​Available Actions

​Media Worker (BullMQ)

scrape_profile

publish_post

apply_watermark

create_teaser

post_scheduler

​Dashboard (React SPA)

​Routes

​AI Chat Widget

​API Proxies

​Data Flow Examples

​Profile Scraping Flow

​Content Classification Flow

​Security Architecture

​Credential Encryption

​Authentication Flow

​RBAC Synchronization

​Nginx Reverse Proxy

​Project Structure

​Performance Considerations

​Next Steps

Quick Start

API Reference

MCP Servers

Deployment

System Architecture

Architecture Overview

Service Layout

MCP Servers (29 Tools Total)

Directus MCP (17 Tools)

Ollama MCP (3 Tools)

Stagehand MCP (9 Tools)

MCP Configuration

Data Layer (Directus Collections)

Action Runner System

How It Works

Available Actions

Media Worker (BullMQ)

Dashboard (React SPA)

Routes

AI Chat Widget

API Proxies

Data Flow Examples

Profile Scraping Flow

Content Classification Flow

Security Architecture

Credential Encryption

Authentication Flow

RBAC Synchronization

Nginx Reverse Proxy

Project Structure

Performance Considerations

Next Steps