Skip to main content
Dot-art illustration of an iron key ring holding many different keys, dots scattering off one edge

How to connect security tools to Claude Code

Add Splunk, SentinelOne, and other security MCP servers to Claude Code: the commands, scopes, read-only setup, and where these connections fall short.

Guide
Updated 11 min read

Summary

To connect a security tool to Claude Code, add the vendor's MCP server with claude mcp add: a remote server over HTTP with OAuth or a bearer token, or a local server as a stdio command, at local, project, or user scope. Then give it a read-only identity and limit what Claude may call with mcp__server__tool permission rules. As of November 2026, Anthropic's connector directory has no listing from Splunk, CrowdStrike, SentinelOne, Microsoft Sentinel, or Google Security Operations, so most security setups use the vendor's own server, and each one reaches a single product.

In Anthropic's connector directory, November 2026
1,128 connectors
Default cap on one MCP tool result in Claude Code
25,000 tokens
Tool output size that triggers a Claude Code warning
10,000 tokens
Minimum Claude Code version for Splunk's OAuth sign-in
2.1.114

Key takeaways

  1. Use the vendor's own MCP server. Anthropic's directory has no Splunk, CrowdStrike, SentinelOne, Sentinel, or Google SecOps connector as of November 2026.
  2. Give Claude Code its own identity with read-only access, such as a dedicated Splunk OAuth client or a SentinelOne service user scoped to one site.
  3. Keep secrets out of .mcp.json with ${VAR} expansion, and let Claude Code keep OAuth client secrets in the system keychain.
  4. Allow read tools by name with mcp__server__tool rules and leave every state-changing tool on a prompt or a deny rule.
  5. Plan for gaps on real cases. Each server reaches one product, and Claude Code moves tool results over 25,000 tokens out to a file by default.

The short version

Claude Code reaches a security tool through that tool's MCP server, so the work is picking the right server, adding it with one command, and limiting what it can do.

  1. Find the vendor's MCP server. Our survey of security MCP servers lists who publishes each one, where it runs, and what it can read or change.
  2. Create an identity for Claude Code in the security tool: its own OAuth client, service user, or token, with read access only.
  3. Add the server with claude mcp add, using HTTP for a hosted server or stdio for a local one.
  4. Write permission rules that allow the read tools you expect Claude to use, and nothing else.
  5. Run /mcp inside Claude Code to sign in and confirm which tools loaded.

Where this falls short on real investigations

The setup in this guide answers questions about one product at a time. Investigation work strains it in four places.

  • One product per server. Following an alert from EDR to identity to the SIEM means several servers, each with its own sign-in, tool names, and failure modes, and Claude has to join the results inside one conversation. When one server's sign-in has expired, that part of the evidence is missing.
  • Result size. Claude Code warns at 10,000 tokens per tool result and moves anything over 25,000 tokens to a file by default. Vendors cap results too: Splunk fails calls that run longer than one minute or return more than 1,000 events, and Sentinel's data lake tools stream for at most 120 seconds. A long lookback on a busy account can exceed these before Claude sees the answer. Cotool's Splunk integration avoids the event cap by calling the Splunk REST API instead: searches run as jobs with a timeout you set from 1 to 10 minutes, return partial results at the timeout, and page through large result sets.
  • Uneven client support. Splunk's OAuth flow covers Splunk Cloud only, Microsoft doesn't document Claude Code for Sentinel MCP, and Google's remote server needs the customer ID, region, and project ID on every request.
  • No loop around the session. Nothing in this setup picks up alerts as they fire, records a verdict in a case, or runs again tomorrow. Someone starts each session and supplies the alert.

Some vendors also meter MCP features into their own assistants: Splunk's saia_ tools require Splunk AI Assistant, and Sentinel's entity analyzer bills Security Compute Units.

Take an MFA push fatigue alert from Okta. In Claude Code, an analyst pastes the alert into a session, Claude searches Splunk for the user's sign-ins within the 1,000-event cap, and the analyst adds an Okta server to see the pushes, then copies the verdict into the case by hand. In Cotool's AI SOC, the alert routes to a triage agent as it arrives. The agent queries Okta and Splunk through its integrations, out of over 100 built-in ones plus custom MCP and CLI tools, and closes or escalates the alert with the evidence attached. A state-changing action such as suspending the user waits for approval in the app or Slack. Cotool also runs an MCP server of its own, so Claude Code can reach those alerts, logs, and detections as one more server; the third setup below adds it.

How Claude Code connects to MCP servers

Claude Code talks to MCP servers over three transports. HTTP is the recommended option for remote servers, SSE is deprecated, and stdio runs the server as a local process on your machine. A remote server can authenticate with OAuth, which Claude Code runs in your browser, or with a static header such as a bearer token.

Where you add a server decides who gets it:

ScopeLoads inShared with teamStored in
Local (default)Current project onlyNo~/.claude.json
ProjectCurrent project onlyYes, through version control.mcp.json in the project root
UserAll your projectsNo~/.claude.json

When the same server name appears in more than one scope, local wins over project, and project over user. Claude Code asks for approval before it uses a server from a project's .mcp.json in an interactive session, and a freshly cloned repository can't approve its own servers. That prompt doesn't appear in claude -p runs or Agent SDK sessions, which load project servers without asking.

Connectors you've added in claude.ai also reach Claude Code in the terminal, VS Code, and JetBrains, where their tools appear as mcp__claude_ai_<server>__<tool>. Remote connectors in the directory are for paid plans only.

Every MCP tool is also a permission target. A rule of mcp__<server>__<tool> matches one tool and mcp__<server>__* matches all tools from a server, in the same allow, ask, and deny lists that govern shell commands and file edits.

What's available for security tools

Anthropic's connector directory is a thin source of security connectors. As of November 2026 it lists 1,128 connectors, and its category filter has no security category. None of the listings come from Splunk, CrowdStrike, SentinelOne, Microsoft Sentinel, Google Security Operations, Wiz, Panther, Okta, or Palo Alto Networks. The security-adjacent entries include:

  • Elastic Cloud Serverless, which runs ES|QL queries and workflows through Kibana's MCP server.
  • Zscaler MCP Server, which lists 402 tools across Zscaler services. Read-only tools are on by default, write tools need explicit allowlisting, and destructive calls need an in-session confirmation token.
  • Rapid7 Bulk Export, which exports InsightVM vulnerability and asset data into a local DuckDB database for SQL queries.
  • Tines, which hosts MCP servers you build from your own Tines workflows.
  • Snyk, 1Password, Vanta, Drata, Datadog, and Sentry, which sit next to security operations rather than inside it.

A directory listing is not a security review. Claude Code's security documentation says Anthropic reviews connectors against its listing criteria but "does not security-audit or manage any MCP server."

For SIEM and EDR platforms, you add the vendor's server yourself. Splunk documents a Claude Code setup and lists Claude Code 2.1.114 or newer as compatible with its OAuth sign-in. SentinelOne's Purple AI MCP README includes a Claude Code command. Panther's Remote MCP is a standard OAuth endpoint at https://api.<your-panther-host>/mcp. Google's remote SecOps server needs Google credentials and an IAM role. Microsoft's Sentinel MCP setup guide documents Visual Studio Code, Security Copilot, Copilot Studio, and Microsoft Foundry, not Claude Code. For vendor-specific detail, see the Splunk MCP server guide and the CrowdStrike MCP guide.

Step by step: three verified setups

All three setups follow the vendors' own documentation as of November 2026.

Splunk Cloud Platform with OAuth

A Splunk admin installs the MCP Server for Splunk platform app (1.2.1 or higher) on Splunk Cloud Platform 10.3.2512.11 or later, enables token authentication, and assigns the mcp_tool_execute capability to the role that will sign in. The admin then creates an OAuth client under Settings, Authentication methods, Splunk OAuth Clients, with the redirect URI http://localhost:8787/callback, and shares the client ID, client secret, and MCP endpoint over a secure channel. Splunk shows the secret once. OAuth is for Splunk Cloud Platform customers only.

On the analyst's workstation, register the server. Claude Code prompts for the client secret and stores it in the system keychain or a credentials file, not in the config:

claude mcp add --transport http \
  --callback-port 8787 \
  --client-id <client-id> \
  --client-secret \
  splunk-mcp <mcp-endpoint>

Splunk's OAuth server advertises more scopes than most clients handle, so pin the scopes. Open ~/.claude.json, find the splunk-mcp entry, and add scopes to its oauth block:

"splunk-mcp": {
  "type": "http",
  "url": "<mcp-endpoint>",
  "oauth": {
    "clientId": "<client-id>",
    "callbackPort": 8787,
    "scopes": "openid offline_access"
  }
}

Start claude, run /mcp, and sign in to Splunk in the browser that opens. Run /mcp again to confirm splunk-mcp is connected and to see its tools. If sign-in fails with redirect_uri does not match, the callback port differs from the one registered in Splunk. On Splunk Enterprise, which has no OAuth option, the app issues encrypted tokens that work only for MCP. Pass one as an Authorization: Bearer header from an environment variable rather than pasting it into the command.

SentinelOne Purple AI MCP as a local server

Purple AI MCP is read-only by design and runs as a local stdio process. In the SentinelOne console, create a service user token under Policy & Settings, User Management, Service Users, with Account or Site permissions rather than Global. Each server instance serves one account or site.

SentinelOne's README gives this one-line command, which is fine for a personal setup but writes the token into ~/.claude.json:

claude mcp add --transport stdio purple-mcp \
  --env PURPLEMCP_CONSOLE_TOKEN=your_token \
  --env PURPLEMCP_CONSOLE_BASE_URL=https://your-console.sentinelone.net \
  -- uvx --from git+https://github.com/Sentinel-One/purple-mcp.git purple-mcp --mode stdio

To share the setup with your team without sharing the token, commit a project .mcp.json that reads both values from each analyst's environment:

{
  "mcpServers": {
    "purple-mcp": {
      "type": "stdio",
      "command": "uvx",
      "args": ["--from", "git+https://github.com/Sentinel-One/purple-mcp.git", "purple-mcp", "--mode", "stdio"],
      "env": {
        "PURPLEMCP_CONSOLE_TOKEN": "${PURPLEMCP_CONSOLE_TOKEN}",
        "PURPLEMCP_CONSOLE_BASE_URL": "${PURPLEMCP_CONSOLE_BASE_URL}"
      }
    }
  }
}

Each analyst exports the two variables, from a secrets manager or a shell profile outside the repository, then approves the server the first time Claude Code asks.

Cotool's MCP server over OAuth

Cotool runs a remote MCP server that signs in with your normal Cotool login, so there is no token to create or store. Add it at user scope with your region's URL (https://eu.cotool.ai/api/mcp for EU organizations):

claude mcp add --transport http --scope user cotool https://app.cotool.ai/api/mcp

Run /mcp, pick cotool, and sign in to Cotool in the browser; the consent screen shows which account and organization Claude Code will act as. From then on, Claude Code acts with that user's role and permissions: it can search and read alerts and change their status, list log tables and run log queries, and run agents. It can also write and test detection code over a past time window without creating alerts. It can't delete anything, and when an integration needs a secret, it returns a link to Cotool's connect form instead of taking the secret in chat. Every tool call is recorded in Cotool's audit log as user.calledMcpTool. The same URL works in Cursor from ~/.cursor/mcp.json, if your team prefers that client; Claude Code vs Cursor for security compares the two.

Lock down what Claude can do

Start from read-only and widen only when a workflow needs it.

  • Read-only identity. Use a role or service user that can't change anything: a Splunk role with only mcp_tool_execute, a SentinelOne Account or Site token, or the --read-only flag on CrowdStrike's self-hosted falcon-mcp, which registers only read tools.
  • Server-side tool controls. Splunk admins can turn tools off for everyone and, from app version 1.3, map tools to roles. Panther's consent screen lets each analyst grant less than their role allows.
  • Pinned OAuth scopes. Claude Code's oauth.scopes field is its supported way to request only the scopes your security team approves.
  • Permission rules. Allow the read tools you expect, by name, in .claude/settings.json. In the default mode Claude Code prompts for anything else, and a deny rule blocks a tool even if something else allows it:
{
  "permissions": {
    "allow": [
      "mcp__splunk-mcp__splunk_run_query",
      "mcp__splunk-mcp__splunk_get_indexes",
      "mcp__purple-mcp__powerquery"
    ],
    "deny": [
      "mcp__falcon-mcp__falcon_execute_rtr_read_only_command"
    ]
  }
}
  • Unattended runs. For claude -p jobs, use the dontAsk permission mode, which denies every call that would otherwise prompt, so only tools on your allow list run.
  • Organization control. Admins can deploy a fixed server set with managed-mcp.json or restrict servers with allowedMcpServers and deniedMcpServers.

These rules live in each engineer's settings and each session. An agent platform applies the same controls to the agent itself: in Cotool, an agent can call only the tools it has been granted, state-changing tools can require approval in the app or Slack, and one click makes an agent read-only.

Common mistakes

  • Pasting tokens into commands or committed files. A header or --env value typed into claude mcp add lands in ~/.claude.json and your shell history. Use ${VAR} references in .mcp.json, and note that Claude Code reads its own and your cloud provider's credential variables as empty in a remote server's URL and headers.
  • Leaving out the type field. A JSON entry with a url but no type is read as a stdio server and skipped.
  • Using an admin or Global token. Purple AI MCP's PowerQuery results come back wrong with a Global token, and an admin token gives Claude more reach than any investigation needs.
  • Exposing a local server on the network. falcon-mcp's HTTP transports have no authentication by default, and Purple AI MCP has none built in. Keep them on localhost.
  • Putting --header before the server name. Splunk notes that --header can swallow the arguments after it, so name the server and endpoint first.
  • Defining one server name in two scopes. Claude Code warns when the same name points at different endpoints, and OAuth sign-ins are stored per endpoint.

Checklist

  • The vendor's MCP server is identified, along with its release status and caps.
  • Claude Code has its own read-only identity in the security tool.
  • The server is added at the right scope: local for personal use, project for a shared setup.
  • No token appears in .mcp.json, the repository, or a pasted command.
  • OAuth scopes are pinned where the server advertises more than you need.
  • /mcp shows the server connected with the tool count you expect.
  • Allow rules name the read tools; state-changing tools prompt or are denied.
  • Unattended claude -p jobs run in dontAsk mode.

FAQ

Is there a Splunk or CrowdStrike connector in Claude's connector directory?
No Splunk or CrowdStrike connector appears in Claude's connector directory as of November 2026, and neither do SentinelOne, Microsoft Sentinel, Google Security Operations, Wiz, Panther, or Okta. Security-adjacent listings include Elastic Cloud Serverless, Zscaler, Rapid7 Bulk Export, Tines, and Snyk. Splunk and CrowdStrike publish their own MCP servers, which you add to Claude Code with claude mcp add.
Can Claude Code take actions in my security tools?
Claude Code can take actions in a security tool only if that tool's MCP server exposes write tools and your permission rules let Claude call them. Splunk's MCP tools and Purple AI MCP only read. In the default permission mode Claude Code asks before the first use of each tool, and a deny rule blocks a tool outright.
Where does Claude Code store MCP server credentials?
Claude Code stores local- and user-scoped MCP servers, including any header values you typed, in ~/.claude.json, and project-scoped servers in .mcp.json in the repository. OAuth client secrets go to the system keychain on macOS or a credentials file, not the config. Reference tokens as ${VAR} in .mcp.json so the file never holds the secret.
Why does Claude Code cut off a large MCP tool result?
Claude Code warns when an MCP tool result passes 10,000 tokens and, by default, saves results over 25,000 tokens to a file instead of the conversation. Raise the cap with the MAX_MCP_OUTPUT_TOKENS environment variable. Vendors set their own caps too: Splunk's MCP server fails any call that returns more than 1,000 events.

Sources

  1. Connect Claude Code to tools via MCP · accessed November 16, 2026
  2. Configure permissions (Claude Code docs) · accessed November 16, 2026
  3. Settings files and precedence (Claude Code docs) · accessed November 16, 2026
  4. Security (Claude Code docs) · accessed November 16, 2026
  5. Connectors and plugins (Claude Marketplace) · accessed November 16, 2026
  6. Claude marketplace sitemap, connector listings · accessed November 16, 2026
  7. Discover tools that work with Claude (Anthropic) · accessed November 16, 2026
  8. Elastic Cloud Serverless connector for Claude · accessed November 16, 2026
  9. Zscaler MCP Server connector for Claude · accessed November 16, 2026
  10. Rapid7 Bulk Export connector for Claude · accessed November 16, 2026
  11. Tines connector for Claude · accessed November 16, 2026
  12. Connect Claude to Splunk MCP Server · accessed November 16, 2026
  13. OAuth for MCP Server (Splunk documentation) · accessed November 16, 2026
  14. Set up the Splunk MCP Server · accessed November 16, 2026
  15. Connecting to the MCP Server and settings (Splunk documentation) · accessed November 16, 2026
  16. MCP Server tools (Splunk documentation) · accessed November 16, 2026
  17. Sentinel-One/purple-mcp on GitHub · accessed November 16, 2026
  18. CrowdStrike/falcon-mcp on GitHub · accessed November 16, 2026
  19. Panther Remote MCP (Beta) · accessed November 16, 2026
  20. Remote MCP server for Google SecOps (google/mcp-security docs) · accessed November 16, 2026
  21. Get started with Microsoft Sentinel MCP server · accessed November 16, 2026
  22. Microsoft Sentinel MCP server pricing, limits, and availability · accessed November 16, 2026
  23. Cotool AI SOC · accessed November 16, 2026
  24. Cotool SIEM · accessed November 16, 2026
  25. MCP Server - Cotool Documentation · accessed November 16, 2026
  26. Connect an App - Cotool Documentation · accessed November 16, 2026
  27. Splunk Integration - Cotool Documentation · accessed November 16, 2026
  28. Tools and Integrations - 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