Elasticsearch 7.x MCP Server

MCP Server for connecting to your Elasticsearch cluster directly from any MCP Client (like Claude Desktop, Cursor).

[!IMPORTANT] This fork targets Elasticsearch 7.x only. It pins the @elastic/elasticsearch 7.17 client, whose product check accepts servers older than 7.14. For an Elasticsearch 8.x cluster, use the upstream project @awesome-ai/elasticsearch-mcp, which this fork is derived from — the 8.x client cannot talk to a 7.x server, and vice versa.

This server connects agents to your Elasticsearch data using the Model Context Protocol. It allows you to interact with your Elasticsearch indices through natural language conversations.

Feature Overview

Tools come in three sets. Only the first is always exposed; the other two are opt-in through an environment variable, so a production deployment can offer diagnostics without offering deletes. Gating happens at registration: a disabled tool never appears in tools/list, so the model cannot call it and it costs nothing in the agent's context.

Always available — read and write data

Cluster

• elasticsearch_health: cluster health, optionally down to index level
• cluster_info: cluster name, Elasticsearch version and build flavour

Index operations

• list_indices: list indices, filtered by an Elasticsearch wildcard (log-*)
• create_index: create an index with optional settings and mappings
• reindex: copy an index, optionally filtered by a query or transformed by a script
• get_aliases: which aliases point at which indices

Mappings

• get_mappings: the fields of an index, as dotted paths with their types, then the raw mapping
• create_mapping: create or update the mapping of an index

Search and data

• search: run a query DSL search, with highlighting injected over every text field — nested ones included — unless the query brings its own highlight
• count: how many documents match, without transferring any
• get_document: fetch one document by id
• bulk: index many documents at once

Templates

• create_index_template: create or update a composable index template
• get_index_template: read index templates

Fields

• field_caps: which fields exist across an index pattern, and whether each is searchable and aggregatable. Takes a wildcard, unlike get_mappings — and it is the only way to see a field mapped as two different types across indices, which makes an aggregation over it silently partial rather than failing.
• analyze: the terms a text is broken into, which is what a query must produce to match. This is the answer to "my search returns nothing and I do not know why".

Tasks

• get_task: progress of a long-running task, such as the one reindex returns

ES_ADMIN_TOOLS=true — diagnostics (read-only)

These only read, so they are safe to enable in production — and are the point of this set: an agent can then explain why an index is unhealthy without anyone logging into the cluster.

• explain_allocation: why a shard is unassigned, with each allocator's decision
• list_shards: shard-level state, leading with the copies that are not STARTED
• list_nodes: heap, CPU, load and disk pressure per node
• get_index_stats: per-index counters — size, segments, indexing, search, merges
• get_index_settings: an index's settings (refresh_interval, replicas, read-only blocks)
• get_cluster_settings: cluster settings that were overridden at runtime
• list_tasks: what the cluster is currently running

ES_ALLOW_DESTRUCTIVE=true — irreversible

Intended for a staging environment, and off by default so production cannot reach them at all.

• delete_index: delete an index and its data
• delete_document: delete one document by id
• delete_by_query: delete every document matching a query — asynchronous, it returns a task id and the deletion continues in the background
• delete_index_template: delete an index template

Even with the flag on, these refuse a wildcard, a comma-separated list, * and _all: they act on one named index at a time. A model that mistakes logs-* for a single index gets a refusal instead of an emptied cluster.

ES_ECS_TOOLS=true — ECS log search (read-only)

Four tools for clusters whose application logs are in ECS. They take named parameters instead of a query DSL and answer in log lines instead of JSON documents, because the schema is known in advance — which is also what lets them request only the fields they print.

They need ES_ECS_INDEX_PATTERN, which has no default: the server refuses to start with the flag on and the pattern missing. A guess like logs-* would sweep whichever indices happen to match on your cluster and answer confidently from the wrong data.

• search_logs: recent events, newest first, one line each — filter by service, levels/minLevel, host, logger, dataset, traceId and free text, over a window given as 15m, 2h, 7d
• log_histogram: counts per time bucket, to see when something started, peaked or stopped. The bucket width is derived from the window, and empty buckets are kept because a gap is part of the answer
• error_summary: errors grouped by error.type, with counts, first and last occurrence, affected services and a sample message
• top_values: the most frequent values of a field — what your cluster actually indexes, before you filter on a guess

[!NOTE] These target ECS 1.x field types, because that is what Elasticsearch 7.8 can express. match_only_text arrived in 7.14 and wildcard in 7.9, so a mapping pushed to a 7.8 cluster necessarily predates ECS 1.12, where error.message and error.stack_trace moved to those types.

The consequence is visible in the tools: error.message is text with no keyword sub-field, so errors cannot be grouped by message — error_summary groups by error.type and gives events lacking one their own reported bucket rather than dropping them. Grouping by error.stack_trace looks possible, since ECS 1.x types it as keyword, but ECS sets ignore_above: 1024 on keywords, so a longer trace is not indexed and would silently vanish from the aggregation.

Enabling this set adds about 8.4 KB to the tools/list response every session, which is why it is a flag: a cluster whose logs are not in ECS would pay for four tool schemas that can only return nothing.

How It Works

1. The MCP Client analyzes your request and determines which Elasticsearch operations are needed.
2. The MCP server carries out these operations (listing indices, fetching mappings, performing searches).
3. The MCP Client processes the results and presents them in a user-friendly format.

Getting Started

Prerequisites

• An Elasticsearch 7.x instance (tested against 7.8; the 7.17 client supports 6.8 through 7.x)
• Elasticsearch credentials — an API key, or a username and password
• An MCP client: Claude Code, Claude Desktop, Codex, Cursor, or anything else that speaks MCP over stdio

Authenticate to GitHub Packages, once

[!IMPORTANT] This package is published to GitHub Packages, not npmjs.com, and GitHub Packages requires a token even for public packages. Until you add one, every install below fails with a 401. Put it in your user-level ~/.npmrc:

@agrica:registry=https://npm.pkg.github.com
//npm.pkg.github.com/:_authToken=YOUR_GITHUB_TOKEN

YOUR_GITHUB_TOKEN is a personal access token with the read:packages scope.

Keep it in your own ~/.npmrc rather than a project file — a token committed to a repository is a leaked token, and some package managers refuse to read one from there at all.

Connect it to your client

Every example below sets ES_HOST and ES_API_KEY. Swap in ES_USERNAME/ES_PASSWORD for basic auth or the ES_OAUTH_* variables for OAuth2 (see OAuth2), add ES_ADMIN_TOOLS=true to get the diagnostic tools, and set ES_INSTANCE_LABEL when more than one instance is declared — see Configuration Options.

<details open> <summary><strong>Claude Code</strong> — <code>claude mcp add</code></summary>

claude mcp add elasticsearch7 \
  --env ES_HOST=https://your-cluster:9200 \
  --env ES_API_KEY=your-api-key \
  --env ES_ADMIN_TOOLS=true \
  -- npx -y @agrica/elasticsearch7-mcp

Then /mcp in a session lists the server and its tools.

Two details that are easy to get wrong:

• Everything after -- is the command that runs the server; without it, Claude Code would try to parse -y as one of its own flags.
• Do not put the server name straight after --env — the CLI reads it as another KEY=value pair and rejects it. Above, the name comes first, which is why it works.

The server is added at local scope, so it loads in the current project only. Add --scope user to get it everywhere, or --scope project to write it into .mcp.json and share it with your team — mind that a committed .mcp.json would carry your API key, so prefer user scope for credentials.

</details>

<details> <summary><strong>Claude Desktop</strong> — <code>claude_desktop_config.json</code></summary>

Edit claude_desktop_config.json — Settings > Developer > Edit Config opens it, or find it at %APPDATA%\Claude\ on Windows and ~/Library/Application Support/Claude/ on macOS:

{
  "mcpServers": {
    "elasticsearch7": {
      "command": "npx",
      "args": ["-y", "@agrica/elasticsearch7-mcp"],
      "env": {
        "ES_HOST": "https://your-cluster:9200",
        "ES_API_KEY": "your-api-key",
        "ES_ADMIN_TOOLS": "true"
      }
    }
  }
}

Restart Claude Desktop afterwards; it only reads that file at startup.

</details>

<details> <summary><strong>Codex</strong> — <code>codex mcp add</code> or <code>config.toml</code></summary>

codex mcp add elasticsearch7 \
  --env ES_HOST=https://your-cluster:9200 \
  --env ES_API_KEY=your-api-key \
  -- npx -y @agrica/elasticsearch7-mcp

Or write it into ~/.codex/config.toml by hand. Note that Codex spells the table mcp_servers with an underscore, and that the environment goes in its own sub-table rather than inline:

[mcp_servers.elasticsearch7]
command = "npx"
args = ["-y", "@agrica/elasticsearch7-mcp"]

[mcp_servers.elasticsearch7.env]
ES_HOST = "https://your-cluster:9200"
ES_API_KEY = "your-api-key"
ES_ADMIN_TOOLS = "true"

/mcp inside Codex confirms the server is loaded.

</details>

<details> <summary><strong>Any other MCP client</strong> — stdio, no port</summary>

The server is a plain stdio MCP server, so anything on the MCP client list works. It needs three things: the command npx, the arguments -y @agrica/elasticsearch7-mcp, and the ES_* variables in its environment. It never listens on a port, and writes nothing but MCP protocol to stdout — diagnostics go to stderr.

</details>

Configuration Options

The Elasticsearch MCP Server supports configuration options to connect to your Elasticsearch:

[!NOTE] You must provide either an API key or both username and password for authentication.

Environment Variable	Description	Required
ES_HOST	Your Elasticsearch instance URL(s) - supports single URL or comma-separated multiple URLs (also supports legacy HOST)	Yes
ES_API_KEY	Elasticsearch API key for authentication (also supports legacy API_KEY)	No
ES_USERNAME	Elasticsearch username for basic authentication (also supports legacy USERNAME)	No
ES_PASSWORD	Elasticsearch password for basic authentication (also supports legacy PASSWORD)	No
ES_CA_CERT	Path to custom CA certificate for Elasticsearch SSL/TLS (also supports legacy CA_CERT)	No
ES_OAUTH_TOKEN_URL	OAuth2 token endpoint. Setting it turns on OAuth2, which then takes precedence over the API key and basic auth. Must be https:// (plain http:// is accepted only for localhost).	No
ES_OAUTH_CLIENT_ID	OAuth2 client id. Required once ES_OAUTH_TOKEN_URL is set.	No
ES_OAUTH_CLIENT_SECRET	OAuth2 client secret. Required once ES_OAUTH_TOKEN_URL is set, unless the _FILE form is used.	No
ES_OAUTH_CLIENT_SECRET_FILE	Path to a file holding the secret, for a mounted Docker secret. Read and trimmed at startup.	No
ES_OAUTH_SCOPE	Scope to request, if the provider needs one.	No
ES_OAUTH_AUDIENCE	Audience to request. Auth0 needs it to issue a JWT; Keycloak and Azure AD use ES_OAUTH_SCOPE instead.	No
ES_OAUTH_AUTH_STYLE	post (default) sends the credentials in the form body; basic sends them in an HTTP Basic header.	No
ES_REQUEST_TIMEOUT	Per-request timeout in milliseconds. Default 30000 — raise it if aggregations over many indices time out.	No
ES_MAX_RETRIES	Retries per request. Default 3; 0 disables them.	No
ES_MAX_RESULT_BYTES	Ceiling on one tool result. Default 32768. Past it, detail is omitted and the result says so.	No
ES_INSTANCE_LABEL	Free-text name of this deployment, e.g. production. Shown as the server title, so several instances declared side by side are distinguishable.	No
ES_ADMIN_TOOLS	true to also expose the read-only diagnostic tools. Default off.	No
ES_ALLOW_DESTRUCTIVE	true to also expose the irreversible tools. Default off.	No
ES_ECS_TOOLS	true to also expose the ECS log search tools. Default off.	No
ES_ECS_INDEX_PATTERN	Index pattern the ECS log tools query, e.g. logs-app-*. No default: with ES_ECS_TOOLS on and this unset, the server refuses to start.	Only with ES_ECS_TOOLS

[!WARNING] ES_ADMIN_TOOLS, ES_ALLOW_DESTRUCTIVE and ES_ECS_TOOLS have no un-prefixed legacy alias, unlike the connection variables above. That is deliberate: a bare ADMIN_TOOLS or ALLOW_DESTRUCTIVE in an environment is far too easy to set by accident for something that decides whether deletes are reachable.

Both accept true or 1; anything else, including an unset variable, means off.

OAuth2

Set ES_OAUTH_TOKEN_URL, ES_OAUTH_CLIENT_ID and ES_OAUTH_CLIENT_SECRET and the server obtains a client_credentials token, renews it before it expires, and sends it as Authorization: Bearer … on every request. It takes precedence over ES_API_KEY and ES_USERNAME/ES_PASSWORD, and says so on stderr at startup if those are still set — so you can see which identity is really talking to the cluster.

What it is for. An Elasticsearch 7.x cluster cannot validate a third-party OAuth2 token itself: the JWT realm arrived in 8.2, and the 7.x OIDC realm needs a platinum licence and a browser. So this is for a cluster reached through a gateway that validates the token and forwards the request. Point ES_HOST at the gateway.

Keep the secret out of the config file. A project .mcp.json is checked into version control, so reference the secret instead of pasting it:

{
  "mcpServers": {
    "elasticsearch": {
      "command": "npx",
      "args": ["-y", "@agrica/elasticsearch7-mcp"],
      "env": {
        "ES_HOST": "https://es-gateway.internal",
        "ES_OAUTH_TOKEN_URL": "https://idp.internal/realms/data/protocol/openid-connect/token",
        "ES_OAUTH_CLIENT_ID": "mcp-elasticsearch",
        "ES_OAUTH_CLIENT_SECRET": "${ES_OAUTH_CLIENT_SECRET}",
        "ES_OAUTH_SCOPE": "es:read"
      }
    }
  }
}

${VAR} is expanded from your own environment. The alternatives are a user-scoped server entry (claude mcp add --scope local, stored outside the repository) or ES_OAUTH_CLIENT_SECRET_FILE pointing at a mounted file.

Ask for the scope you can use. If the deployment runs without ES_ALLOW_DESTRUCTIVE, request a read-only scope: a write scope buys nothing when no write tool is registered, and widens what the secret is worth if it leaks. When the gateway refuses a request for want of a scope, the error names the one it asked for.

Result size

A tool result is capped at 32 KB (ES_MAX_RESULT_BYTES). This matters on a logging cluster: before the cap, one list_shards call over a year of daily indices returned 385 KB — around 96 000 tokens — in a single answer, which is more than most sessions can hold.

When a result is trimmed it says so, says how much went, and says how to ask a smaller question. Three tools shape their answers around it:

• list_indices and list_shards return a readable summary; the same rows as text are behind verbose.
• search caps size at 100 per call and tells you the from to page with.
• get_mappings lists the fields first and the raw mapping second, so a thousand-field index still answers the question it was asked.

Four tools — list_indices, list_shards, get_index_settings and get_mappings — also return their answer as typed structured output, so a client can read the rows instead of parsing the text. It is assembled from whatever room the readable answer left, and reports returned against total so a partial listing is visible as a number.

Run pnpm run measure against the built output to see the current figures for your own configuration.

Labelling several instances

Most setups declare this server more than once — one entry per cluster. The entries are otherwise identical, so a client shows two servers with the same name and nothing to tell them apart. ES_INSTANCE_LABEL becomes the server's display title, and it is the natural place to say which environment an entry reaches:

{
  "mcpServers": {
    "es7-prod": {
      "command": "npx",
      "args": ["-y", "@agrica/elasticsearch7-mcp"],
      "env": {
        "ES_HOST": "https://es-prod:9200",
        "ES_API_KEY": "prod-key",
        "ES_INSTANCE_LABEL": "production",
        "ES_ADMIN_TOOLS": "true"
      }
    },
    "es7-staging": {
      "command": "npx",
      "args": ["-y", "@agrica/elasticsearch7-mcp"],
      "env": {
        "ES_HOST": "https://es-staging:9200",
        "ES_API_KEY": "staging-key",
        "ES_INSTANCE_LABEL": "staging",
        "ES_ADMIN_TOOLS": "true",
        "ES_ALLOW_DESTRUCTIVE": "true"
      }
    }
  }
}

That pair is the intended shape: diagnostics on both, deletes only on staging. Production keeps the tools that explain an unhealthy index and never exposes one that can remove data — the model cannot call what was never registered.

The label is also printed to stderr at startup, which is where to look when a client reports a connection but you cannot tell which cluster answered.

Multiple URLs Configuration

You can configure multiple Elasticsearch nodes for high availability and load balancing:

{
  "mcpServers": {
    "elasticsearch7-mcp": {
      "command": "npx",
      "args": [
        "-y",
        "@agrica/elasticsearch7-mcp"
      ],
      "env": {
        "ES_HOST": "https://es-node1:9200,https://es-node2:9200,https://es-node3:9200",
        "ES_API_KEY": "your-api-key"
      }
    }
  }
}

The client will automatically handle failover and load balancing between the configured nodes.

Running with Docker

Each release publishes a multi-arch image (linux/amd64, linux/arm64) to the GitHub Container Registry:

docker pull ghcr.io/agrica/elasticsearch7-mcp:latest

The server speaks stdio, so the container needs an interactive stdin and no published port. In an MCP client:

{
  "mcpServers": {
    "elasticsearch7-mcp": {
      "command": "docker",
      "args": [
        "run", "--rm", "-i",
        "-e", "ES_HOST",
        "-e", "ES_API_KEY",
        "ghcr.io/agrica/elasticsearch7-mcp:latest"
      ],
      "env": {
        "ES_HOST": "your-elasticsearch-host",
        "ES_API_KEY": "your-api-key"
      }
    }
  }
}

[!NOTE] Like the npm package, the image lives in GitHub Packages: pulling it requires a token with the read:packages scope, even though the repository is public.

The image needs no port published and no volume: it speaks stdio, and the MCP client owns its stdin and stdout.

Example Queries

[!TIP] Here are some natural language queries you can try with your MCP Client.

Cluster Management

• "What is the health status of my Elasticsearch cluster?"
• "How many active nodes are in my cluster?"

Index Operations

• "What indices do I have in my Elasticsearch cluster?"
• "Create a new index called 'users' with 3 shards and 1 replica."
• "Reindex data from 'old_index' to 'new_index'."

Mapping Management

• "Show me the field mappings for the 'products' index."
• "Add a keyword type field called 'tags' to the 'products' index."

Search & Data Operations

• "Find all orders over $500 from last month."
• "Which products received the most 5-star reviews?"
• "Bulk import these customer records into the 'customers' index."

Template Management

• "Create an index template for logs with pattern 'logs-*'."
• "Show me all my index templates."

Diagnostics (needs ES_ADMIN_TOOLS=true)

• "The 'logs-2026' index is yellow — why are its shards unassigned?"
• "Is any node close to a disk watermark?"
• "Which of my indices is the largest, and how much of it is deleted documents?"
• "Has anyone disabled shard allocation on this cluster?"
• "Is a reindex still running?"

ECS logs (needs ES_ECS_TOOLS=true)

• "What errors has the billing service logged in the last hour?"
• "When did the 5xx spike start, and is it still going?"
• "Which hosts are producing these timeouts?"
• "What log levels does this cluster actually index?"

Destructive (needs ES_ALLOW_DESTRUCTIVE=true)

• "Delete the 'smoke-test-source' index."
• "Remove every document older than 2024 from 'logs-archive'."

Troubleshooting

Symptom	Cause
npm error code E401 on install or npx	No GitHub Packages token in your user-level ~/.npmrc. See Authenticate to GitHub Packages.
Server error: ... invalid url at startup	ES_HOST is unset or malformed. It is validated at startup on purpose, rather than failing later on the first query.
Server error: ... ES_OAUTH_CLIENT_SECRET ... is required	An ES_OAUTH_* variable is set but the block is incomplete. A half-configured factor is refused rather than falling back to another identity.
Error: Authentication failed before any request was sent: … on every tool	The token endpoint could not be reached, or refused the credentials. The message carries the provider's own error field. Nothing was sent to the cluster.
Error: … error="insufficient_scope", and asks for scope "…"	The gateway wants a scope the token does not carry. Add it to ES_OAUTH_SCOPE.
The client connects, but a diagnostic or delete tool is missing	That set is gated. Set ES_ADMIN_TOOLS=true or ES_ALLOW_DESTRUCTIVE=true and restart the client.
Server error: ... ES_ECS_INDEX_PATTERN is required	ES_ECS_TOOLS is on with no pattern. There is deliberately no default — set it to the pattern holding your ECS logs.
search_logs returns nothing, with no error	A keyword filter did not match the indexed spelling. service.name, log.level and the rest are exact and case-sensitive; run top_values on the field to see the real values.
top_values refuses a field as "analysed text"	Aggregations need a keyword. Try <field>.keyword, or ask field_caps which fields the pattern reports as aggregatable.
Refusing to act on the pattern "logs-*"	Working as intended: destructive tools take one concrete index name, never a pattern, even with the flag on.
A connection error mentioning the product check	The cluster is 8.x, or unreachable. This build talks to 7.x only.

Found a bug or want a tool that is missing? Open an issue on the GitHub repository. To work on the code, start from CONTRIBUTING.md.