AnythingLLM MCP Server (Fixed Fork)

This is a fixed fork of raqueljezweb/anythingllm-mcp-server Original package has broken API endpoints. This fork fixes them.

An MCP (Model Context Protocol) server that enables seamless integration between AnythingLLM and MCP-compatible clients like Claude Code, Claude Desktop, and more.

Quick Start for Claude Code

1. Add to your MCP configuration

Add this to ~/.claude.json (or %USERPROFILE%\.claude.json on Windows):

{
  "mcpServers": {
    "anythingllm": {
      "type": "stdio",
      "command": "node",
      "args": ["C:/path/to/anythingllm-mcp-server-fork/src/index.js"],
      "env": {
        "ANYTHINGLLM_API_KEY": "YOUR-API-KEY-HERE",
        "ANYTHINGLLM_BASE_URL": "http://localhost:3001"
      }
    }
  }
}

2. Get your API key

1. Open AnythingLLM → Settings → API Keys
2. Create a new API key
3. Copy it to the config above

3. Restart Claude Code

The MCP server will auto-initialize with the env variables.

---

For LLMs: How to Use This MCP Server

IMPORTANT: This section is for AI assistants using this MCP server.

Initialization (Required First!)

Before any operation, you MUST initialize the client:

mcp__anythingllm__initialize_anythingllm
  apiKey: "YOUR-API-KEY"
  baseUrl: "http://localhost:3001"

If env variables are set, initialization happens automatically.

Core Workflow: RAG (Retrieval-Augmented Generation)

Step 1: Create a workspace

mcp__anythingllm__create_workspace
  name: "my-knowledge-base"

Step 2: Add documents (embed text)

mcp__anythingllm__embed_text
  slug: "my-knowledge-base"
  texts: ["Document content here...", "Another document..."]

Step 3: Query with RAG

mcp__anythingllm__chat_with_workspace
  slug: "my-knowledge-base"
  message: "What does the documentation say about X?"
  mode: "query"   # Use "query" for RAG, "chat" for conversation

All Available Tools

Tool	Purpose	Key Parameters
initialize_anythingllm	REQUIRED FIRST	apiKey, baseUrl
list_workspaces	List all workspaces	-
get_workspace	Get workspace details + documents	slug
create_workspace	Create new workspace	name
update_workspace	Update workspace settings	slug, updates
delete_workspace	Delete workspace	slug
embed_text	Add text to RAG	slug, texts[]
embed_webpage	Scrape & add webpage	slug, url
list_documents	List docs in workspace	slug
delete_document	Remove document	slug, documentId
chat_with_workspace	Query RAG / Chat	slug, message, mode
search_workspace	Vector similarity search	slug, query
get_chat_history	Get conversation history	slug

Common Patterns

Pattern 1: Scrape web docs → Add to RAG → Query

# 1. Create workspace
create_workspace(name="docs")

# 2. For each page, scrape with Crawl4AI then embed
#    (Use mcp__crawl4ai__md to get markdown, then embed_text)
embed_text(slug="docs", texts=[markdown_content])

# 3. Query
chat_with_workspace(slug="docs", message="How do I...", mode="query")

Pattern 2: Delete a document by docId

# 1. List documents to get docId
list_documents(slug="my-workspace")
# Returns: documents[].docId like "79d25253-a860-4c30-95ab-48d13dd4fd04"

# 2. Delete using docId
delete_document(slug="my-workspace", documentId="79d25253-a860-4c30-95ab-48d13dd4fd04")

Pattern 3: Clear and recreate workspace

# AnythingLLM doesn't have "clear all docs" - delete and recreate
delete_workspace(slug="old-workspace")
create_workspace(name="old-workspace")  # Fresh start

Important Notes for LLMs

1. Always initialize first - Every new session needs initialize_anythingllm
2. Use mode: "query" for RAG - This retrieves relevant documents. mode: "chat" doesn't use RAG.
3. list_documents returns docId - Use this UUID for delete_document
4. No clear chat history API - Workaround: delete and recreate workspace
5. Slugs are auto-generated - When you create "My Workspace", slug becomes "my-workspace"

---

Bugs Fixed in This Fork

Bug	Original Behavior	Fixed Behavior
list_documents	Returns empty array []	Returns actual documents
delete_document	Doesn't delete (wrong API)	Works with docId/filename/docpath
delete_workspace	JSON parse error on "OK"	Returns {success: true}
workspace.workspace	Assumed object	Handles both array and object
Multiple endpoints	Wrong paths	Correct AnythingLLM v1 API paths

See Issue #1 on original repo.

---

Installation Options

Option 1: Clone this fork (Recommended)

git clone https://github.com/Tapiocapioca/anythingllm-mcp-server.git
cd anythingllm-mcp-server
npm install

Option 2: NPM (Original - HAS BUGS)

# NOT RECOMMENDED - original package has broken endpoints
npm install -g anythingllm-mcp-server

---

Configuration Examples

Claude Code (~/.claude.json)

{
  "mcpServers": {
    "anythingllm": {
      "type": "stdio",
      "command": "node",
      "args": ["C:/AI/TestVari/anythingllm-mcp-server-fork/src/index.js"],
      "env": {
        "ANYTHINGLLM_API_KEY": "XXXXX-XXXXXX-XXXXXX-XXXXXXX",
        "ANYTHINGLLM_BASE_URL": "http://localhost:3001"
      }
    }
  }
}

Claude Desktop

macOS: ~/Library/Application Support/Claude/claude_desktop_config.json Windows: %APPDATA%\Claude\claude_desktop_config.json

{
  "mcpServers": {
    "anythingllm": {
      "command": "node",
      "args": ["/path/to/anythingllm-mcp-server-fork/src/index.js"],
      "env": {
        "ANYTHINGLLM_API_KEY": "your-key",
        "ANYTHINGLLM_BASE_URL": "http://localhost:3001"
      }
    }
  }
}

---

Troubleshooting

"Client not initialized" error

→ Call initialize_anythingllm with your API key first

"Workspace not found"

→ Use list_workspaces to see available slugs

Documents not appearing

→ Use get_workspace instead of list_documents (more reliable)

Connection refused

→ Ensure AnythingLLM is running on port 3001

---

Security

• Never commit API keys to version control
• Use environment variables for credentials
• API keys can be regenerated in AnythingLLM settings

---

Credits

• Original: raqueljezweb/anythingllm-mcp-server
• Fixes: Tapiocapioca/anythingllm-mcp-server
• AnythingLLM by Mintplex Labs
• Model Context Protocol by Anthropic