Skip to main content
Dot-art illustration of an electrical plug with a coiled cord, dots scattering off one edge

Security MCP servers: what the official ones can do

Who publishes each security MCP server, where it runs, how it signs in, and what it can read or change: Splunk, CrowdStrike, Sentinel, Google SecOps, and more.

Research brief
Updated 9 min read

Summary

As of November 2026, most large security platforms publish an official MCP server, and Microsoft Sentinel, Google Security Operations, CrowdStrike, Wiz, and Panther host theirs and sign in with the platform's own identity and roles. The gaps are in scope and limits rather than availability: each server reaches one vendor's data, several expose only read tools or are still in beta or early access, and caps such as Splunk's 1,000-event response limit and Sentinel's 120-second streaming limit cut long investigations short.

In the Google SecOps remote MCP server reference
70 tools
Read-only triage tools in Microsoft Sentinel MCP
27 tools
Splunk MCP response size that fails a call
1,000 events
Sentinel data lake MCP streaming limit
120 seconds

Key findings

  1. Microsoft Sentinel, Google SecOps, CrowdStrike, and Wiz run their MCP servers on their own infrastructure. Splunk's runs on your search head, and falcon-mcp, Purple AI MCP, and Cortex run on your machine or container.
  2. Write access varies widely. Google SecOps and Okta expose case, rule, and user changes, while Splunk, Sentinel triage, Purple AI MCP, and Cortex built-ins only read.
  3. Every server reaches one vendor. An investigation across EDR, SIEM, and identity means running several servers, each with its own sign-in and limits.
  4. Check the caps before building on a server. Splunk fails responses over 1,000 events, and Sentinel streams for at most 120 seconds.
  5. Several releases aren't GA. Panther Remote MCP is in open beta, Okta's managed server is in early access, and SentinelOne calls Purple AI MCP unofficial.

What we checked

We read the documentation and repositories for eleven MCP servers from security vendors, as of November 2026: Splunk, CrowdStrike (both the hosted CrowdStrike MCP and the self-hosted falcon-mcp), Microsoft Sentinel, Google Security Operations, SentinelOne, Panther, Wiz, Elastic, Okta, and Palo Alto Networks Cortex. For each server we recorded who publishes it, where it runs, how it authenticates, what it can read or change, and its release status. The table adds Cotool's own server, described on the same axes from its documentation. Tool counts are our own counts of each vendor's published tool list. We didn't run the servers against live tenants, so this covers what each vendor documents, not how well the tools perform on real cases.

The servers side by side

ServerRuns whereSign-inReads and actsStatus
Splunk MCP ServerApp on your Splunk search headMCP-only token; OAuth on Splunk CloudAbout 34 tools; search and read onlySplunkbase app, docs at version 2.0
CrowdStrike MCPHosted in Falcon (US-1, US-2, EU-1)Falcon roles and permissionsTools found through search_tools and execute_toolGA since September 2026
falcon-mcpLocal process or your containerFalcon API client ID and secretModule-based tools, with a --read-only flagMaintenance mode
Microsoft Sentinel MCPHosted by MicrosoftMicrosoft Entra IDLake queries, entity analysis, 27 read-only triage toolsLive; some pages marked prerelease
Google SecOps remote MCPHosted, regional Chronicle endpointsGoogle credentials plus an IAM role70 tools, including case, rule, and SOAR changesListed without a preview label
Purple AI MCP (SentinelOne)Local, or self-hosted without built-in authConsole token for one account or siteRead-only alerts, PowerQuery, vulnerabilities, threat intelOpen source; "not an official" product
Panther Remote MCPHosted at your api. subdomainPer-user OAuth 2.1 with PKCEAlerts, detections, data lake, within your roleOpen beta
Wiz MCPHosted by WizService account or OAuthSecurity Graph, Wiz agent analyses, AI SkillsGA since July 2026
Elastic Agent Builder MCPYour Kibana endpointAPI key; OAuth 2.1 on ServerlessBuilt-in and custom Agent Builder toolsGA on Serverless and 9.3
Okta MCP ServerSelf-hosted, or Okta-managed as a paid add-onDevice grant or private key JWTFull read and write, confirms destructive callsSelf-hosted GA; managed in early access
Cortex MCP ServerDownloaded to a machine or containerNot publishedBuilt-in tools read; custom tools can writeReleased December 2025
Cotool MCP serverHosted by Cotool, US and EU regionsCotool login over OAuth; user's own roleAgents, alerts, logs, detections; no deletesOn for every organization

Two vendors changed course in 2026. CrowdStrike put falcon-mcp in maintenance mode, with only security and dependency fixes from now on, and recommends the hosted CrowdStrike MCP it made generally available in September 2026. Elastic deprecated its standalone mcp-server-elasticsearch repository in favor of the MCP endpoint built into Elastic Agent Builder. Google still publishes local servers in google/mcp-security, but its own README marks the fully managed remote server as recommended.

Where they fall short for investigation work

The limits show up once a question crosses tools or needs a lot of data, which is most alerts an agent would be asked to work.

  • One vendor per server. Each of the eleven vendor servers reaches only its own product's data, so tracing an alert from EDR to identity to cloud logs means connecting several servers, each with its own sign-in, scopes, and failure modes. The agent, or the analyst driving it, has to join the results.
  • Read-only surfaces. Splunk's tools search and read but don't change anything, Sentinel's 27 triage tools are all reads, Purple AI MCP is read-only by design, and Cortex's built-in tools don't write to the tenant. An agent on these servers can gather evidence but can't close the alert or contain a host.
  • Hard caps. Splunk fails a call that runs longer than one minute or returns more than 1,000 events. Sentinel's data lake tools stream for at most 120 seconds with an 800-character query window, and its entity analyzer allows 200 runs an hour and 500 a day per tenant. A lookback over a busy user's activity can hit these caps before it answers. Cotool's Splunk integration, which uses the Splunk REST API rather than the MCP server, runs search jobs with a timeout you set from 1 to 10 minutes, returns partial results when a search reaches it, and pages through large result sets.
  • Setup friction. Google's remote server needs the customer ID, region, and project ID on every request, and a Google developer advocate writes that the struggle to get clients to pass them "continues." Sentinel's triage tools work only in your home tenant and can't choose a workspace. Purple AI MCP serves one account or site per server.
  • Release status. Panther Remote MCP is in open beta, Okta's managed server is in early access with general availability planned for early 2027, and Sentinel's graph tools are in preview.
  • Unauthenticated defaults. The falcon-mcp README warns that its HTTP transports have no authentication by default, and Purple AI MCP has no built-in authentication, so neither should be exposed on a network as shipped.

Take a suspicious OAuth grant for one user. With direct connections, an analyst's client needs the identity provider's server for the user's sign-ins and a separate connection for the file sharing the grant touched, each signed in separately with its own caps, and the analyst joins the results. In Cotool, an investigation agent hands 30 days of the user's Okta sign-ins and the Google Drive sharing activity to investigator subagents that run in parallel, a reviewer subagent checks the verdict, and the alert is closed or escalated with the evidence attached.

What the hosted servers provide

The hosted servers use the platform's existing identity, and several log each call to the platform's audit trail. Microsoft's Sentinel MCP server is fully hosted, needs no infrastructure, signs in with Entra, and is offered at no extra cost, though the data lake queries it runs are billed. Google's remote SecOps server exposes 70 tools spanning search, cases, detection rules, parsers, feeds, and SOAR actions, governed by IAM, organization policy, and audit logs. CrowdStrike MCP has nothing to install, follows Falcon roles, and records activity in the Falcon audit log.

In Panther Remote MCP, each analyst connects with their own OAuth sign-in, the consent screen lets them grant less than their role allows, a reduced role makes the next token refresh fail closed, and every call lands in Panther's audit log as an AI_TOOL_INVOKE event. Splunk lets admins turn individual tools on or off and map them to roles, and its tokens work only for MCP. Okta's server asks the user to confirm destructive operations before it runs them.

What vendors sell next to their servers

Most of these vendors also sell an AI assistant or agents inside their own console, and some MCP features depend on those products. CrowdStrike announced its managed MCP server in the same release as new Charlotte AI AgentWorks features. Splunk's saia_ tools require Splunk AI Assistant. Microsoft bills Sentinel's entity analyzer in Security Compute Units, the capacity unit Security Copilot uses, and one of its MCP collections exists to build Security Copilot agents. Palo Alto Networks describes the Cortex MCP Server as complementing its native Cortex Agentic Assistant, Purple AI MCP's first tool sends questions to SentinelOne's Purple AI (see built-in security AI vs AI SOC platforms for how these console assistants compare with cross-stack agents), and Okta's managed server needs its own subscription. Price an MCP server together with whatever it calls into.

Connecting a server to an AI client

Most of these servers work with any MCP client. Claude Code and Cursor both take local and remote servers but differ in approvals and admin controls, as Claude Code vs Cursor for security lays out. Our guide to connecting security tools to Claude Code walks through scopes, permissions, and three verified setups. For single-vendor detail, see the Splunk MCP server guide and the CrowdStrike MCP guide.

Where an agent platform fits

Connecting one server per vendor still leaves someone to join the results into one investigation. An agent platform does that joining itself: Cotool's AI SOC has over 100 built-in integrations plus custom MCP and CLI tools, and its agents gather context across those tools for each alert. Each integration is authenticated once for the organization, with OAuth or an API key, and an agent can call only the tools it has been granted. State-changing actions can require approval in the app or in Slack, and one click makes an agent read-only.

Cotool also runs its own hosted MCP server, with the same identity model as the hosted vendor servers. A user adds one URL to Claude Code, Cursor, or another client and signs in with their Cotool login over OAuth; there are no API keys, and the app acts with that user's role and permissions. It reaches agents, alerts, logs, and detections, including writing and testing detection code against a past window of your data without creating alerts. Integration secrets never pass through the chat, and every tool call lands in Cotool's audit log as user.calledMcpTool.

FAQ

Which security vendors have an official MCP server?
As of November 2026, Splunk, CrowdStrike, Microsoft Sentinel, Google Security Operations, Panther, Wiz, Elastic, Okta, and Palo Alto Networks Cortex all publish an official MCP server. SentinelOne maintains Purple AI MCP on GitHub but says it is not an official SentinelOne product. Tines takes a different route: you build your own MCP server out of Tines workflows.
Are security MCP servers read-only?
Some security MCP servers are read-only and some are not. Splunk's tools search and read, Microsoft Sentinel's triage tools only read, Purple AI MCP is read-only by design, and Cortex's built-in tools don't write to the tenant. Google SecOps's remote server and Okta's server include tools that update cases, create rules, or change users.
Do security MCP servers work with Claude Code?
Most security MCP servers work with Claude Code because they use standard MCP over HTTP or stdio. Splunk lists Claude Code 2.1.114 or newer as compatible with its OAuth sign-in, and SentinelOne's README includes a Claude Code command. Microsoft's Sentinel MCP setup guide documents Visual Studio Code, Security Copilot, Copilot Studio, and Microsoft Foundry rather than Claude Code.
What is the biggest risk with self-hosted security MCP servers?
The biggest risk with self-hosted security MCP servers is exposing one on a network without authentication. The falcon-mcp README warns that its HTTP transports have no authentication by default, and Purple AI MCP has no built-in authentication at all. Keep these servers bound to localhost or put them behind an authenticating proxy.

Sources

  1. MCP Server tools (Splunk documentation) · accessed November 16, 2026
  2. Set up the Splunk MCP Server · accessed November 16, 2026
  3. OAuth for MCP Server (Splunk documentation) · accessed November 16, 2026
  4. Connecting to the MCP Server and settings (Splunk documentation) · accessed November 16, 2026
  5. CrowdStrike/falcon-mcp on GitHub · accessed November 16, 2026
  6. Falcon MCP modules overview · accessed November 16, 2026
  7. CrowdStrike Delivers the Next Evolution of the Agentic SOC · accessed November 16, 2026
  8. What is Microsoft Sentinel's support for MCP? · accessed November 16, 2026
  9. Get started with Microsoft Sentinel MCP server · accessed November 16, 2026
  10. Microsoft Sentinel MCP triage tool collection · accessed November 16, 2026
  11. Microsoft Sentinel MCP server pricing, limits, and availability · accessed November 16, 2026
  12. google/mcp-security on GitHub · accessed November 16, 2026
  13. Remote MCP server for Google SecOps (google/mcp-security docs) · accessed November 16, 2026
  14. MCP Reference: chronicle.googleapis.com · accessed November 16, 2026
  15. Google Cloud MCP servers, supported products · accessed November 16, 2026
  16. Google Cloud remote MCP server for SecOps (Google Cloud Community) · accessed November 16, 2026
  17. Sentinel-One/purple-mcp on GitHub · accessed November 16, 2026
  18. Panther Remote MCP (Beta) · accessed November 16, 2026
  19. panther-labs/mcp-panther on GitHub · accessed November 16, 2026
  20. Introducing Wiz MCP: Build AI Security Agents with Wiz · accessed November 16, 2026
  21. Add security context to operational investigations with AWS DevOps Agent and Wiz · accessed November 16, 2026
  22. Elastic Agent Builder MCP server · accessed November 16, 2026
  23. elastic/mcp-server-elasticsearch on GitHub · accessed November 16, 2026
  24. okta/okta-mcp-server on GitHub · accessed November 16, 2026
  25. Okta Managed MCP Server (Okta Help Center) · accessed November 16, 2026
  26. Okta MCP Server use cases (Okta blog) · accessed November 16, 2026
  27. Cortex MCP Server (Cortex XSIAM documentation) · accessed November 16, 2026
  28. Use the Cortex MCP server · accessed November 16, 2026
  29. Introducing the Cortex MCP Server (Palo Alto Networks blog) · accessed November 16, 2026
  30. Introducing MCP servers in Tines · accessed November 16, 2026
  31. Cotool AI SOC · accessed November 16, 2026
  32. Splunk Integration - Cotool Documentation · accessed November 16, 2026
  33. Tools and Integrations - Cotool Documentation · accessed November 16, 2026
  34. MCP Server - Cotool Documentation · accessed November 16, 2026

Keep reading

Attackers are scaling with tokens

Cotool helps defenders operate at machine speed. See how security teams are scaling Detection & Response.

Book a demo