AnythingLLM Server

A Model Context Protocol (MCP) server that lets MCP-compatible clients (VS Code, Claude Desktop, Cursor, and others) interact with AnythingLLM. Available tools:

• Workspace management
• Chat and thread operations
• Document operations and embeddings
• Vector search
• System and model inspection

Requirements

• Python 3.10+ — verify: python --version or python3 --version. Download if not installed.
• uv (Python package manager) — verify: uv --version. Install instructions.
• AnythingLLM running and reachable — open http://localhost:3001 in a browser to confirm. If it runs in Docker, make sure the port is published (e.g., -p 3001:3001).
• AnythingLLM API key — create one in the AnythingLLM web interface (look for API Keys or Developer API in Settings). Copy the key and store it securely; it is shown only once.

Installation

git clone https://github.com/andreperez/anythingllm-mcp.git
cd anythingllm-mcp
uv sync

uv sync reads pyproject.toml and installs all required dependencies into a local virtual environment. You do not need to create or activate the environment manually — uv run handles that automatically.

Configuration

The server reads configuration from these environment variables at startup:

• ANYTHINGLLM_BASE_URL: Base URL of your AnythingLLM instance. Defaults to http://localhost:3001 if not set.
• ANYTHINGLLM_API_KEY: API key for authenticating with AnythingLLM. This value is required.

The server does not load .env files by itself. You must either:

• Set the variables in the shell before starting the server.
• Use a tool that loads .env files for you.
• Pass the variables directly in your MCP client configuration.

Option 1: Set variables in your current shell

PowerShell:

$env:ANYTHINGLLM_BASE_URL = "http://localhost:3001"
$env:ANYTHINGLLM_API_KEY = "your_api_key_here"

Bash:

export ANYTHINGLLM_BASE_URL="http://localhost:3001"
export ANYTHINGLLM_API_KEY="your_api_key_here"

Command Prompt (cmd.exe):

set ANYTHINGLLM_BASE_URL=http://localhost:3001
set ANYTHINGLLM_API_KEY=your_api_key_here

Use the host-exposed URL for your deployment. For Docker port mappings like 3002:3001, set ANYTHINGLLM_BASE_URL to http://localhost:3002 because the MCP server connects from the host, not from inside the container.

Option 2: Use a .env file with uv

Copy the example file:

PowerShell:

Copy-Item .env.example .env

Bash:

cp .env.example .env

Command Prompt (cmd.exe):

copy .env.example .env

Edit .env so it contains your real values:

ANYTHINGLLM_BASE_URL=http://localhost:3001
ANYTHINGLLM_API_KEY=replace_with_your_real_api_key

Then start the server with:

uv run --env-file .env anythingllm-mcp

This tells uv to load variables from .env before starting the server.

Option 3: Pass variables in your MCP client configuration

Most MCP clients accept an env block in their configuration file (see the Client setup examples below). When the env block contains ANYTHINGLLM_API_KEY and ANYTHINGLLM_BASE_URL, you do not need to export variables in the shell or use a .env file.

Client setup

Replace /path/to/anythingllm-mcp in the examples below with the absolute path where you cloned the repository. Examples:

• Linux / macOS: /home/youruser/anythingllm-mcp
• Windows: C:\\Users\\youruser\\anythingllm-mcp (use double backslashes \\ inside JSON strings)

VS Code

Add to your user or workspace mcp.json. Open the Command Palette (Ctrl+Shift+P on Windows/Linux, Cmd+Shift+P on macOS) and run MCP: Open User Configuration:

{
  "servers": {
    "anythingllm": {
      "type": "stdio",
      "command": "uv",
      "args": [
        "run",
        "--directory",
        "/path/to/anythingllm-mcp",
        "anythingllm-mcp"
      ],
      "env": {
        "ANYTHINGLLM_API_KEY": "${input:anythingllm_api_key}",
        "ANYTHINGLLM_BASE_URL": "http://localhost:3001"
      }
    }
  }
}

VS Code supports ${input:name} to prompt for the API key on each session.

In a multi-root workspace, prefer an absolute path for --directory if the MCP server lives outside the current folder root.

If VS Code shows only a partial tool list after you update the server or switch from another AnythingLLM integration to this local one:

1. Run MCP: List Servers and confirm anythingllm or your chosen server name starts successfully.
2. Run MCP: Reset Cached Tools to clear stale tool metadata.
3. Restart the MCP server from MCP: List Servers.

Claude Desktop

Edit claude_desktop_config.json (Settings → Developer → Edit Config):

{
  "mcpServers": {
    "anythingllm": {
      "command": "uv",
      "args": [
        "run",
        "--directory",
        "/path/to/anythingllm-mcp",
        "anythingllm-mcp"
      ],
      "env": {
        "ANYTHINGLLM_API_KEY": "YOUR_API_KEY",
        "ANYTHINGLLM_BASE_URL": "http://localhost:3001"
      }
    }
  }
}

Cursor

Add to .cursor/mcp.json in your project root, or to the global config at ~/.cursor/mcp.json (Linux/macOS) / %USERPROFILE%\.cursor\mcp.json (Windows):

{
  "mcpServers": {
    "anythingllm": {
      "command": "uv",
      "args": [
        "run",
        "--directory",
        "/path/to/anythingllm-mcp",
        "anythingllm-mcp"
      ],
      "env": {
        "ANYTHINGLLM_API_KEY": "YOUR_API_KEY",
        "ANYTHINGLLM_BASE_URL": "http://localhost:3001"
      }
    }
  }
}

Note: ANYTHINGLLM_BASE_URL defaults to http://localhost:3001, which is the application port inside the container. If you publish Docker as 3002:3001, use http://localhost:3002 in client configuration.

Run (standalone)

If you already exported the variables in your shell:

uv run anythingllm-mcp

If you want to load them from .env at launch time:

uv run --env-file .env anythingllm-mcp

Development

Install test dependencies and run the test suite:

uv sync --extra dev
uv run pytest

Troubleshooting

"API key is missing" or 401 Unauthorized

The server could not authenticate with AnythingLLM. Check that:

1. ANYTHINGLLM_API_KEY is set in your environment or in the env block of your MCP client configuration.
2. The key matches an active key in AnythingLLM (Settings → API Keys).
3. The key has no leading or trailing whitespace.

"Connection refused" or timeout

The server could not reach AnythingLLM. Check that:

1. AnythingLLM is running (open http://localhost:3001 in a browser).
2. ANYTHINGLLM_BASE_URL points to the correct host and port.
3. No firewall or VPN is blocking the connection.

Docker port mismatch

If AnythingLLM runs in Docker with a port mapping like 3002:3001:

• Inside the container, the application listens on port 3001.
• On the host, you access it on port 3002.

Because the MCP server runs on the host (not inside the container), set:

ANYTHINGLLM_BASE_URL=http://localhost:3002

Always use the host port (the left side of the -p or ports: mapping).

MCP server running inside a container (Dev Containers, Docker Compose)

If the MCP server itself runs inside a container (for example, a VS Code Dev Container), localhost refers to that container's own network — not to the host machine. To reach AnythingLLM running on the host or in another container with a published port, use host.docker.internal:

ANYTHINGLLM_BASE_URL=http://host.docker.internal:3001

If both the MCP server and AnythingLLM are containers on the same Docker network (for example, in the same docker-compose.yml), use the service name as hostname instead:

# "anythingllm" is the service name defined in docker-compose.yml
ANYTHINGLLM_BASE_URL=http://anythingllm:3001

In this case, use the container port (3001), not the host-published port.

host.docker.internal is supported on Docker Desktop (Windows and macOS) and on Docker Engine 20.10+ for Linux (requires --add-host=host.docker.internal:host-gateway or the equivalent extra_hosts in Compose).

Security notes

• Never commit real API keys.
• Use environment variables for secrets.
• Use least-privileged AnythingLLM API tokens when possible.

License

MIT