Elasticsearch/OpenSearch MCP Server

Overview

A Model Context Protocol (MCP) server implementation that provides Elasticsearch and OpenSearch interaction. This server enables searching documents, analyzing indices, and managing cluster through a set of tools.

Demo

https://github.com/user-attachments/assets/f7409e31-fac4-4321-9c94-b0ff2ea7ff15

Features

General Operations

• general_api_request: Perform a general HTTP API request. Use this tool for any Elasticsearch/OpenSearch API that does not have a dedicated tool.

Index Operations

• list_indices: List all indices.
• get_index: Returns information (mappings, settings, aliases) about one or more indices.
• create_index: Create a new index.
• delete_index: Delete an index.
• create_data_stream: Create a new data stream (requires matching index template).
• get_data_stream: Get information about one or more data streams.
• delete_data_stream: Delete one or more data streams and their backing indices.

Document Operations

• search_documents: Search for documents.
• index_document: Creates or updates a document in the index.
• get_document: Get a document by ID.
• delete_document: Delete a document by ID.
• delete_by_query: Deletes documents matching the provided query.

Cluster Operations

• get_cluster_health: Returns basic information about the health of the cluster.
• get_cluster_stats: Returns high-level overview of cluster statistics.

Alias Operations

• list_aliases: List all aliases.
• get_alias: Get alias information for a specific index.
• put_alias: Create or update an alias for a specific index.
• delete_alias: Delete an alias for a specific index.

Analyzer Operations

• analyze_text: Analyze text using a specified analyzer or custom analysis chain. Useful for debugging search queries and understanding how text is tokenized.

Configure Environment Variables

The MCP server supports the following environment variables:

Basic Authentication (Username/Password)

• ELASTICSEARCH_USERNAME: Username for basic authentication
• ELASTICSEARCH_PASSWORD: Password for basic authentication
• OPENSEARCH_USERNAME: Username for OpenSearch basic authentication
• OPENSEARCH_PASSWORD: Password for OpenSearch basic authentication

API Key Authentication (Elasticsearch only) - Recommended

• ELASTICSEARCH_API_KEY: API key for Elasticsearch or Elastic Cloud Authentication.

Connection Settings

• ELASTICSEARCH_HOSTS / OPENSEARCH_HOSTS: Comma-separated list of hosts (default: https://localhost:9200)
• ELASTICSEARCH_CLUSTERS / OPENSEARCH_CLUSTERS: Inline JSON object for named cluster configurations. When set, tools can target a specific cluster with the optional cluster parameter.
• ELASTICSEARCH_CLUSTERS_FILE / OPENSEARCH_CLUSTERS_FILE: Path to a JSON file with the clusters object. Recommended when the configuration is embedded inside another JSON file (e.g. the MCP client config) because it avoids JSON-in-JSON escaping. Takes precedence over the inline variable when both are set.
• DEFAULT_CLUSTER: Default cluster name to use when multi-cluster configuration is set and a tool call omits cluster (defaults to the first configured cluster).
• VERIFY_CERTS: Whether to verify SSL certificates (default: false)
• REQUEST_TIMEOUT: Request timeout in seconds (optional, uses client default if not set)

Multiple Cluster Configuration

By default, the server uses a single Elasticsearch cluster from ELASTICSEARCH_HOSTS, ELASTICSEARCH_USERNAME, ELASTICSEARCH_PASSWORD, and ELASTICSEARCH_API_KEY, or a single OpenSearch cluster from OPENSEARCH_HOSTS, OPENSEARCH_USERNAME, and OPENSEARCH_PASSWORD. To configure multiple named clusters, set ELASTICSEARCH_CLUSTERS (or OPENSEARCH_CLUSTERS) to a JSON object inside the MCP server configuration. Because the value is a JSON string embedded in another JSON file, the inner quotes need to be escaped:

{
  "mcpServers": {
    "elasticsearch-mcp-server": {
      "command": "uvx",
      "args": [
        "elasticsearch-mcp-server"
      ],
      "env": {
        "ELASTICSEARCH_CLUSTERS": "{\"prod\": {\"hosts\": [\"https://prod-es:9200\"], \"api_key\": \"<PROD_API_KEY>\", \"verify_certs\": true}, \"staging\": {\"hosts\": [\"https://staging-es:9200\"], \"username\": \"elastic\", \"password\": \"<STAGING_PASSWORD>\"}}",
        "DEFAULT_CLUSTER": "prod"
      }
    }
  }
}

For better readability, point ELASTICSEARCH_CLUSTERS_FILE (or OPENSEARCH_CLUSTERS_FILE) at a standalone JSON file instead. The value is just a path so it avoids the JSON-in-JSON escaping:

{
  "mcpServers": {
    "elasticsearch-mcp-server": {
      "command": "uvx",
      "args": [
        "elasticsearch-mcp-server"
      ],
      "env": {
        "ELASTICSEARCH_CLUSTERS_FILE": "/etc/mcp/es-clusters.json",
        "DEFAULT_CLUSTER": "prod"
      }
    }
  }
}

/etc/mcp/es-clusters.json:

{
  "prod": {
    "hosts": ["https://prod-es:9200"],
    "api_key": "<PROD_API_KEY>",
    "verify_certs": true
  },
  "staging": {
    "hosts": ["https://staging-es:9200"],
    "username": "elastic",
    "password": "<STAGING_PASSWORD>"
  }
}

Every tool accepts an optional cluster parameter. If omitted, the server uses DEFAULT_CLUSTER. When DEFAULT_CLUSTER is not set, the first cluster in the JSON object is used as the default. A tool call targeting a specific cluster looks like:

{
  "cluster": "staging",
  "index": "logs-*",
  "body": {
    "query": {
      "match_all": {}
    }
  }
}

MCP Server Authentication (HTTP Transports Only)

When running the MCP server with HTTP-based transports (SSE or Streamable HTTP), you can enable Bearer token authentication to protect the server from unauthorized access.

• MCP_API_KEY: API key for MCP server authentication. Clients must include Authorization: Bearer <MCP_API_KEY> header.

Important Security Notes:

• Authentication is only applicable for HTTP transports (sse, streamable-http). The stdio transport uses local process communication and doesn't require authentication.
• If MCP_API_KEY is not set, the MCP server will be accessible without authentication. This is a security risk when exposing the server over a network.
• For production deployments with HTTP transports, always set MCP_API_KEY.

# Generate a secure API key (example using openssl)
export MCP_API_KEY=$(openssl rand -base64 32)

# Or set a custom API key
export MCP_API_KEY="your-secure-api-key-here"

Disable High-Risk Operations

• DISABLE_HIGH_RISK_OPERATIONS: Set to true to disable all write operations (default: false)
• DISABLE_OPERATIONS: Comma-separated list of specific operations to disable (optional, uses default write operations list if not set)

When DISABLE_HIGH_RISK_OPERATIONS is set to true, all MCP tools that perform write operations are completely hidden from the MCP client. In this mode, the following MCP tools are disabled by default.

• Index Operations:

  • create_index
  • delete_index

• Document Operations:

  • index_document
  • delete_document
  • delete_by_query

• Data Stream Operations:

  • create_data_stream
  • delete_data_stream

• Alias Operations:

  • put_alias
  • delete_alias

• General API Operations:

  • general_api_request

Optionally, you can specify a comma-separated list of operations to disable in the DISABLE_OPERATIONS environment variable.

# Disable High-Risk Operations
export DISABLE_HIGH_RISK_OPERATIONS=true
# Disable specific operations only
export DISABLE_OPERATIONS="delete_index,delete_document,delete_by_query"

GCF Response Encoding (optional)

Opt in to serialize tool-result payloads as GCF (Graph Compact Format), a token-optimized wire format, in the content block the model reads. Elasticsearch returns large, uniform record sets (search hits, aggregation buckets, mappings), the shape GCF compacts best: on representative responses it is ~39% fewer tokens than compact JSON (40% on search hits), losslessly.

export RESPONSE_FORMAT=gcf

structuredContent is preserved unchanged, so a tool's declared output schema still validates and any non-model client keeps receiving JSON; only the model-facing text block is re-encoded. Encoding is fail-safe: any error, including a value outside GCF's canonical int64 numeric domain (which GCF rejects rather than silently approximating), leaves the original JSON result untouched, so a tool call is never dropped over encoding. Default behavior is unchanged when RESPONSE_FORMAT is unset.

Reproduce the token comparison: uv run --with tiktoken python benchmarks/gcf_benchmark.py.

Start Elasticsearch/OpenSearch Cluster

Start the Elasticsearch/OpenSearch cluster using Docker Compose:

# For Elasticsearch
docker-compose -f docker-compose-elasticsearch.yml up -d

# For OpenSearch
docker-compose -f docker-compose-opensearch.yml up -d

The default Elasticsearch username is elastic and password is test123. The default OpenSearch username is admin and password is admin.

You can access Kibana/OpenSearch Dashboards from http://localhost:5601.

Stdio

Option 1: Using uvx