Splunk MCP server: setup, tools, and limits
Splunk's official MCP server runs on your search head and gives Claude Code or Cursor SPL search and read-only tools. Setup, auth, tools, and limits.
Summary
Splunk's official MCP server is the Splunk MCP Server app on Splunkbase, a Splunk-supported app that runs on your search head and lets MCP clients such as Claude Code and Cursor run SPL searches, list indexes and knowledge objects, and read alerts with the calling user's Splunk permissions. It has been generally available since version 1.0.0, and version 2.1.0 is current as of November 2026. Its built-in tools only read, a search that runs longer than a minute or returns more than 1,000 events fails by design, and it sees only what that Splunk deployment can search.
- Current Splunk MCP Server app version, as of November 2026
- 2.1.0
- Maximum run time for a splunk_run_query search
- 1 minute
- Maximum events splunk_run_query returns before it fails
- 1,000 events
Key takeaways
- Install Splunk's own app (Splunkbase app 7931); splunk/splunk-mcp-server2 on GitHub describes itself as unofficial.
- On Splunk Cloud Platform, connect with OAuth and restrict scopes to openid offline_access; on Splunk Enterprise, use an encrypted token per user.
- Give MCP users a dedicated role with mcp_tool_execute and map that role only to the tools they need, because agents act with the user's permissions.
- If investigations need searches longer than a minute or larger than 1,000 events, create a workload pool for splunk_run_query before rollout.
- Plan where writes happen, because the built-in alert, REST, and Enterprise Security tools are read-only.
The short version
Splunk's supported MCP server is the Splunk MCP Server app. It installs on a search head or search head cluster, speaks streamable HTTP, and exposes SPL search, index and metadata discovery, knowledge objects, saved searches, and alert configuration as MCP tools. Splunk says version 2.0 comes preinstalled for eligible customers starting with Splunk Cloud 10.6.
What it gives an agent is read access through quick searches. splunk_run_query stops at one minute or 1,000 events, the alert and Enterprise Security tools are read-only, and OAuth sign-in is limited to Splunk Cloud Platform. That fits an analyst asking quick questions from a chat client. It doesn't fit an agent working an investigation that needs larger searches, evidence from other tools, or a way to act on what it finds.
Limits for investigation work
Splunk designed the MCP server for an analyst asking quick questions of Splunk from a chat client, under the permissions and audit trail they already have. The limits show up when an agent tries to run an investigation end to end.
- Search size.
splunk_run_queryfails when a search runs longer than one minute or returns more than 1,000 events. The documented workaround is a workload pool, which turns the call into a detached, asynchronous search job that the admin has to set up first. Cotool's Splunk integration, which calls the Splunk REST API with a token for a dedicated Splunk user, handles the same problem differently: it runs ad-hoc search jobs with a timeout you set from 1 to 10 minutes (2 by default), returns the results found so far when a search reaches it, and lets the agent page through the rest of a completed job. - Read-only by design. Splunk describes the alert tools as read-only ("an agent can read your alert estate… but it cannot change it"), the
restcommand as read-only, and the Enterprise Security tools as a curated read-only set. No built-in tool updates a finding, closes an investigation, or runs a response action; any write path is a custom tool you build and govern. In Cotool's integration, the one tool that changes Splunk issplunk_createDetection, which creates a saved search; an agent can call only the tools it has been granted, and state-changing actions can be set to wait for approval in the app or Slack. - Uneven availability. OAuth requires Splunk Cloud Platform,
es_get_ai_analysisis unavailable on-premises, and thesaia_discovery tool depends on Agent Mode in Splunk AI Assistant. - Client plumbing. Clients that can't speak streamable HTTP rely on Node.js and
mcp-remote, which Splunk says are dependencies it doesn't own. - Rate limits per node. Rate-limit counting happens per node and isn't replicated across the stack, so limits on a search head cluster behave per member.
- Splunk only. The server sees what that Splunk deployment has indexed or can reach through federated search. Evidence in an EDR, identity provider, or cloud console needs its own server, and those vary a lot in quality, as the CrowdStrike MCP guide and the survey of official security MCP servers show.
Splunk's own AI sits on the other side of that line. The saia_ tools and Agent Mode come from Splunk AI Assistant, and Enterprise Security's triage and response agents run inside ES. The MCP server gives outside agents a read-only view of the same data; acting on it stays inside Splunk's own products.
What Splunk ships, and what it doesn't
Search results for "Splunk MCP" mix three different things, so it helps to separate them.
- The Splunk MCP Server app (official). Splunkbase app 7931, created by Splunk LLC and marked Splunk Supported. Version 2.1.0 is the default as of November 2026, and Splunkbase lists compatibility with Splunk Enterprise and Splunk Cloud on platform versions 9.3 through 10.7. Splunk declared the server generally available at version 1.0.0, while noting that individual tools inside it can still ship as preview or beta.
- The deprecated cloud endpoint. Before 1.0.0, Splunk also offered an "On-Cloud" (SCS) hosted endpoint. Splunk deprecated it with 1.0.0; it still works for migration but gets no updates.
- Community and sample servers. Repositories such as
splunk/splunk-mcp-server2run locally over stdio, SSE, or Docker and call the Splunk REST API with your credentials. That repository describes itself as "Unofficial" and has a single commit. Treat projects like it as examples, not as the supported path.
The official app runs inside Splunk rather than on the analyst's laptop. Tools execute as the user who called them, so index and knowledge-object permissions apply unchanged, and Splunk writes a provenance label for MCP searches into _audit.
Authentication and roles
The app adds two capabilities. mcp_tool_execute lets a role use MCP tools, and mcp_tool_admin covers tool management and token creation. Since version 1.3, admins can also map specific roles to specific tools, so a user only sees and runs the tools their role is mapped to.
Clients authenticate in one of two ways:
- Encrypted tokens work on every deployment. You generate them in the MCP Server app, they are shown once, and they cannot be used for direct Splunk API calls. Creating one for yourself needs
edit_tokens_ownplusmcp_tool_admin; creating one for any user needsedit_tokens_allplusmcp_tool_admin. Upgrading from 0.2.6 to 1.0.0 was a breaking change because only encrypted tokens work from 1.0.0 on. - OAuth 2.1 with PKCE is available only to Splunk Cloud Platform customers. It needs Splunk Cloud Platform 10.3.2512.11 or later and app version 1.2.1 or later, and Splunk turns it on by default from Splunk Cloud Platform 10.5.2506.8. Splunk lists Claude Code (2.1.114 or newer) and Cursor as compatible, Claude Desktop as depending on configuration, and Claude.ai, VS Code, and Codex as not compatible.
The tools it exposes
Tools are namespaced by source. The core splunk_ tools are on by default:
| Tool group | Tools | What they do |
|---|---|---|
| Search | splunk_run_query, splunk_run_saved_search | Run SPL or SPL2 (with the @spl2 prefix), or an existing saved search |
| Discovery | splunk_get_indexes, splunk_get_index_info, splunk_get_metadata, splunk_get_sourcetypes | Find which indexes, hosts, sources, and sourcetypes hold data |
| Environment | splunk_get_info, splunk_get_user_info, splunk_get_user_list, splunk_get_kv_store_collections, splunk_get_knowledge_objects | Instance details, users, KV Store stats, and 17 knowledge object types |
| Alerts | splunk_list_alerts, splunk_get_alert_details, splunk_list_fired_alerts, splunk_get_fired_alert_details, splunk_get_alert_throttle, splunk_list_active_throttles | Read alert configuration, fired alerts, and suppression state |
Two optional groups depend on other products. The saia_ tools (saia_generate_spl, saia_explain_spl, saia_optimize_spl, saia_ask_splunk_question, and saia_find_data_source) need the Splunk AI Assistant app, and saia_find_data_source also needs its Agent Mode. The o11y_ tools reach Splunk Observability Cloud for APM services, traces, metrics, and SignalFlow. If you run Enterprise Security 8.7 or later with app version 1.3.x or later, a curated set of read-only es_ tools adds investigations, findings, artifacts, notes, response plans, risk scores, and queues, without overriding ES role-based access. For how the assistant behind the saia_ tools compares with other platforms' AI and with cross-stack agents, see built-in security AI vs AI SOC platforms.
Teams can also bring their own tools. Splunk's 2.0 announcement says any Splunk app can register MCP tools backed by an SPL template or a REST call routed through splunkd.
Step by step
These steps follow Splunk's 2.0 documentation. For the general pattern across other security tools, see connecting security tools to Claude Code.
- Clear the network path. On Splunk Cloud Platform, add each MCP client's source IP to the
search-uiallowlist for the Splunk Web endpoint, or tosearch-apiif a client uses the direct management-port endpoint. On every deployment, turn on token authentication. - Install the app. Install the Splunk MCP Server app from Splunkbase on your search head or search head cluster (skip this on Splunk Cloud 10.6 stacks where it is preinstalled). Restart if Splunk prompts you to.
- Create a role. Add
mcp_tool_executeto a dedicated role, and in the app's Tools tab, map that role to the tools it should have. Keepmcp_tool_adminfor administrators. - Set guardrails. In the Guardrails tab, set the tool timeout, default row limit, and rate limits. Splunk sets no rate limits by default.
- Issue credentials. For OAuth (Splunk Cloud Platform), an admin creates a client under Settings, Authentication methods, Splunk OAuth Clients, registers the redirect URI, and shares the client ID, client secret, callback port, and endpoint. For tokens, each user generates an encrypted token in the app, which also displays the endpoint and a sample configuration.
- Connect the client. Use one of the configurations below, then confirm the tools load.
Claude Code with OAuth. Register the server, paste the client secret when prompted, and keep the callback port identical to the redirect URI the admin registered:
claude mcp add --transport http \
--callback-port <callback-port> \
--client-id <client-id> \
--client-secret \
splunk-mcp <mcp-endpoint>
Claude Code has no flag for OAuth scopes, so Splunk says to edit the splunk-mcp entry in ~/.claude.json and add them:
"splunk-mcp": {
"type": "http",
"url": "<mcp-endpoint>",
"oauth": {
"clientId": "<client-id>",
"callbackPort": <callback-port>,
"scopes": "openid offline_access"
}
}
Then start claude, run /mcp, sign in to Splunk in the browser, and run /mcp again to confirm splunk-mcp is connected.
Claude Code with an encrypted token. Claude Code's documented syntax for a remote HTTP server with a bearer header works with the endpoint and token from the Splunk app:
claude mcp add --transport http splunk-mcp https://<MCP_SERVER_ENDPOINT> \
--header "Authorization: Bearer <YOUR_ENCRYPTED_TOKEN>"
Claude Desktop with an encrypted token. Splunk's sample configuration uses mcp-remote to bridge to the HTTP endpoint. Add it to claude_desktop_config.json (Settings, Developer, Edit Config), then fully quit and restart Claude Desktop:
{
"mcpServers": {
"splunk-mcp-server": {
"command": "npx",
"args": [
"-y",
"mcp-remote",
"https://<MCP_SERVER_ENDPOINT>",
"--header",
"Authorization: Bearer <YOUR_ENCRYPTED_TOKEN>"
]
}
}
}
Cursor with OAuth. Register cursor://anysphere.cursor-mcp/oauth/callback as the redirect URI on the Splunk OAuth client, then in Cursor open Settings, Tools and MCPs, New MCP Server:
{
"mcpServers": {
"splunk-mcp-server": {
"url": "<MCP endpoint>",
"auth": {
"CLIENT_ID": "<Client ID>",
"CLIENT_SECRET": "<Client Secret>",
"scopes": ["openid", "offline_access"]
}
}
}
}
On Splunk Enterprise, where OAuth isn't available, use the token-based mcp-remote configuration in Cursor instead.
Connecting Splunk to an agent platform instead
An agent platform connects to Splunk once for the organization rather than per analyst. Cotool's Splunk integration skips the MCP Server app and calls the Splunk REST API:
- Create a token. In Splunk Web, open Settings, Tokens, New Token. Set User to a dedicated Splunk user with search access to only the indexes agents need, not an admin, enter
Cotoolas the Audience, and set an expiration (an empty one expires after 30 days). - Connect. In Cotool, open Integrations, Splunk, Connect, and enter the base URL with the management port (usually
https://<host>:8089), the token without aBearerprefix, and an optional query timeout from 1 to 10 minutes. - Review the map. Cotool maps your indexes, sourcetypes, key fields, and sample rows on connect and weekly after that, and agents use the map to pick where to search.
The integration's tools list datasets and sourcetype fields, run and page through search jobs, list Splunk apps and saved searches, and create saved searches. Cotool rejects searches over index=* and leading-wildcard searches, and every search runs as the token's Splunk user, so that user's roles still decide what agents can see.
Common mistakes
- Installing a community server by accident. Several GitHub and registry listings use "Splunk MCP" in the name. Check for Splunkbase app 7931 and the Splunk Supported label.
- Leaving scopes unrestricted. The Splunk OAuth server advertises more scopes than most clients handle. Without
openid offline_access, Claude Code reportsQuery argument scope is required. - Mismatched callback ports. The port in the client configuration must match the redirect URI registered in Splunk exactly, or sign-in fails with
redirect_uri does not match. - Pasting tokens with whitespace. Claude Code warns about hidden leading or trailing whitespace in headers but uses the value as written, so a token copied with a trailing newline fails.
- Forgetting the allowlist. On Splunk Cloud Platform, a client whose IP isn't on
search-ui(orsearch-apifor the direct endpoint) never reaches the server. - Rotating keys casually. Invalidate Keys revokes every encrypted token at once, because the app uses one key pair for all tokens.
- Disabling TLS checks in production.
ssl_verify = falseinmcp.confandNODE_TLS_REJECT_UNAUTHORIZED=0are for testing with self-signed certificates only.
Where an agent platform fits
A single MCP server answers questions about one tool; an investigation usually spans several. Agent platforms such as Cotool's AI SOC connect to over 100 built-in integrations plus custom MCP and CLI tools, including your own MCP server, and investigate alerts from any source whether or not Cotool stores the logs. That makes Splunk one input among several rather than the whole workflow.
Take an impossible-travel alert from Splunk. Through the Splunk MCP server, an agent can confirm the sign-ins, as long as the search fits in a minute and 1,000 events. The questions that settle the alert sit outside Splunk: whether the EDR saw a new process on the user's laptop, and whether the identity provider shows an MFA push the user approved. In Cotool, the triage agent queries Splunk and those tools directly through its integrations, closes or escalates the alert with that evidence attached, and asks for approval before an action such as isolating the host.
Checklist
- Splunkbase app 7931 installed on the search head or cluster, or confirmed preinstalled on Splunk Cloud 10.6
- Token authentication on; client IPs on the
search-uiorsearch-apiallowlist (Splunk Cloud Platform) - Dedicated role with
mcp_tool_execute, mapped only to needed tools mcp_tool_adminlimited to administrators- Timeout, default row limit, and global and per-tool rate limits set
- OAuth clients created per tool or team (Splunk Cloud Platform), scopes restricted to
openid offline_access - Encrypted tokens issued per user with an expiration (Splunk Enterprise)
- Workload pool created if agents need longer or larger searches
- A saved search in
_auditthat tracks MCP activity - A decision on where findings get updated and response actions run, since the built-in tools only read: a custom tool you build, or an agent platform where state-changing actions wait for approval
FAQ
- Is there an official Splunk MCP server?
- Yes, the official Splunk MCP server is the Splunk MCP Server app, built by Splunk LLC, listed on Splunkbase as app 7931, and marked Splunk Supported. It has been generally available since version 1.0.0. GitHub projects such as splunk/splunk-mcp-server2 are separate, and that one describes itself as unofficial.
- Does the Splunk MCP server work with Splunk Enterprise on-premises?
- The Splunk MCP server works with Splunk Enterprise as well as Splunk Cloud Platform; Splunkbase lists compatibility with platform versions 9.3 through 10.7. On-premises users authenticate with encrypted tokens, because OAuth for the Splunk MCP server requires Splunk Cloud Platform. The Enterprise Security tool es_get_ai_analysis is also not available on on-premises deployments.
- Can the Splunk MCP server change anything in Splunk?
- The Splunk MCP server's built-in tools are designed to read: Splunk documents its alert tools, its Enterprise Security tools, and the rest command as read-only, and splunk_run_query refuses searches it considers unsafe or destructive. Teams can register their own tools backed by an SPL template or a REST call through splunkd, so any write path has to be one you build and govern yourself.
- How do you limit what an AI agent can do through the Splunk MCP server?
- The Splunk MCP server gives admins several controls: the mcp_tool_execute capability, role-to-tool mappings since version 1.3, server-level tool on and off switches, guardrails for timeouts and row limits, and global and per-tool rate limits. Tools run as the calling user, so existing index permissions still apply, and searches carry a provenance label into the _audit index.
Sources
- About MCP Server for Splunk Platform · accessed November 16, 2026
- Set up the Splunk MCP Server (2.0) · accessed November 16, 2026
- MCP Server tools (2.0) · accessed November 16, 2026
- Connecting to the MCP Server and settings (2.0) · accessed November 16, 2026
- OAuth for MCP Server (2.0) · accessed November 16, 2026
- Connect Claude to Splunk MCP Server · accessed November 16, 2026
- Connect Cursor to Splunk MCP Server · accessed November 16, 2026
- MCP Server rate limiting · accessed November 16, 2026
- MCP Server REST command support · accessed November 16, 2026
- Use the Splunk MCP Server app to access tools in Splunk Enterprise Security · accessed November 16, 2026
- Splunk MCP Server on Splunkbase · accessed November 16, 2026
- Splunk MCP Server 2.0 (Splunk blog) · accessed November 16, 2026
- splunk/splunk-mcp-server2 on GitHub · accessed November 16, 2026
- Connect Claude Code to tools via MCP · accessed November 16, 2026
- Connect to local MCP servers - Model Context Protocol · accessed November 16, 2026
- AI SOC Without the Black Box - Cotool · accessed November 16, 2026
- Splunk Integration - Cotool Documentation · accessed November 16, 2026
- Tools and Integrations - Cotool Documentation · accessed November 16, 2026





