MCP Servers

Model Context Protocol servers that extend the AI agent with 29 tools across data, inference, and browser automation

Genie Helper uses three MCP (Model Context Protocol) servers to provide the AI agent with 29 specialized tools for data management, LLM inference, and browser automation.

Architecture

MCP servers run as child processes of AnythingLLM and communicate via stdio transport. They auto-boot on AnythingLLM startup via a patched server/utils/boot/index.js → bootMCPServers() call.

Configuration: storage/plugins/anythingllm_mcp_servers.json

---

Directus MCP Server (17 tools)

Script: scripts/directus-mcp-server.mjs

Provides CRUD operations for Directus collections, user management, file handling, and flow triggering.

Collections

Tool	Description
list-collections	List all custom Directus collections (excludes system tables)
get-collection-schema	Get field schema for a collection with types and requirements

Items CRUD

Tool	Description
read-items	Read items with optional filtering, sorting, pagination, and field selection
read-item	Read a single item by ID (UUID)
create-item	Create a new item with tier-gating for media_jobs operations
update-item	Update an existing item by ID
delete-item	Delete an item from a collection
search-items	Full-text search across items in a collection

Users

Tool	Description
get-me	Get the current authenticated user with full profile fields
list-users	List users with optional filtering (admin operation)
get-user	Get a single user by ID with role information
update-user	Update a user by ID (admin operation)
create-user	Create a new user (requires email, password, role UUID)

Files

Tool	Description
list-files	List files with optional folder filtering and metadata
get-file	Get file metadata and download URL by ID

Flows

Tool	Description
list-flows	List all available Directus Flows with trigger information
trigger-flow	Trigger a flow by UUID with optional JSON payload

Tier Gating

The create-item tool includes built-in rate limiting for media_jobs creation:

• Checks feature access flags (ai_clip_ops, watermark_ops, etc.) against subscription tier
• Enforces usage quotas by counting user_usage_counts within billing cycle
• Throws tier_limit: prefixed errors visible to the agent
• Fails open on unexpected errors to maintain availability

---

Ollama MCP Server (3 tools)

Script: scripts/ollama-mcp-server.mjs

Exposes Ollama’s local LLM inference API for text generation, chat, and model management.

Tool	Description
list-models	List all locally available Ollama models with size and modified timestamps
generate	Single-turn completion with prompt, optional system message, temperature, and max_tokens
chat	Multi-turn chat using messages array with role/content objects

Configuration

• Default URL: http://127.0.0.1:11434
• Default model: qwen-2.5:latest (configurable via OLLAMA_MODEL env var)
• Temperature range: 0-2 (default 0.7)
• Stream mode: Disabled (all responses are synchronous)

---

Stagehand MCP Server (9 tools)

Script: scripts/stagehand-mcp-server.mjs

Wraps the Stagehand browser automation server (port 3002) for web scraping, platform interactions, and HITL (human-in-the-loop) workflows.

Session Management

Tool	Description
start-session	Start a new browser session with configurable headless mode, returns session ID
close-session	End and clean up a browser session

Navigation & Interaction

Tool	Description
navigate	Navigate to a URL in an existing session
act	Perform browser actions using natural language (“Click the Login button”)
observe	Observe interactive elements on the current page

Data Extraction

Tool	Description
extract	Extract structured data using natural language instruction and optional JSON schema
screenshot	Take a screenshot of the current browser page (returns base64 or URL)

Cookie Management

Tool	Description
get-cookies	Get all cookies from the current browser session
set-cookies	Inject cookies for bypassing login (HITL workflow)

Configuration

• Stagehand URL: http://127.0.0.1:3002
• Default model: ollama/qwen-2.5 (for vision/element detection)
• Launch args: Stealth mode enabled (--disable-blink-features=AutomationControlled)
• Window size: 1920x1080
• Resource limits: ~300MB RAM per active session, ~33 concurrent sessions ceiling

Error Handling

All tools surface HTTP errors to the agent with status code and error message extraction from Stagehand responses.

---

Related Collections

• platform_sessions: Stores encrypted browser cookies captured via extension or HITL
• hitl_sessions: Human-in-the-loop login requests (triggers yellow dashboard alert)
• media_jobs: BullMQ job records for scrape, watermark, teaser, publish operations
• action_flows: Flow definitions consumed by Action Runner (slug → steps JSON)

See Also

• Action Runner — Flow-based execution system that uses MCP tools
• Ollama Models — Available LLM models and their roles
• Taxonomy System — 6-concept content classification