Cortex可观测分析与响应MCP服务器

cortex-mcp

An MCP server that lets an AI client run Cortex observable analysis and response, end to end.

Website: https://lidless.dev/cortex-mcp

cortex-mcp is a Model Context Protocol (MCP) server for Cortex, the observable analysis and active-response engine from StrangeBee/TheHive Project. It exists because analysts already drive Cortex by hand through its web UI or raw REST API, and an AI client can do that work faster: detonate an indicator across every applicable analyzer, aggregate the taxonomy verdicts, and pull artifacts without anyone clicking through a dozen jobs. It differs from a generic HTTP bridge by exposing Cortex's real domain model as 30 typed MCP tools, auto-detecting observable data types, fanning out analysis with a cap, and gating every destructive action (responders, deletes, file reads) behind explicit confirmation.

What it does

cortex-mcp connects an MCP-capable AI client (Claude Desktop, Claude Code, Codex CLI, OpenClaw, Hermes, and others) to a running Cortex instance so the model can perform real observable analysis and threat-intelligence enrichment. Cortex is the analyzer/responder engine in the StrangeBee and TheHive SOAR stack: it runs analyzers against observables (IPs, domains, URLs, file hashes, emails, files) and executes responders against TheHive entities. This server speaks Cortex's REST API and projects the full pipeline as MCP tools, resources, and prompts, so an agent can browse analyzer definitions, enable and configure them, submit observables, wait for job reports, extract IOC artifacts, and triage alerts. Auto-detection classifies an observable's data type before analysis, bulk analysis fans out across applicable analyzers and aggregates the taxonomy results, and superadmin tools cover organization and user/API-key management. The result is conversational observable analysis: ask "what does Cortex think of 185.220.101.42?" and get an aggregated multi-analyzer verdict back.

Prerequisites

• Node.js 20 or later
• A running Cortex instance (v3.x recommended)
• A Cortex API key with appropriate permissions

Installation from source

If you prefer to run from a checkout instead of npx:

git clone https://github.com/lidless-labs/cortex-mcp.git
cd cortex-mcp
npm install
npm run build

Then point your client at the built binary (see the per-client recipes below).

Try it (copy-paste MCP client config)

Add this to your MCP client config (this example is Claude Desktop's claude_desktop_config.json; the same command/args/env shape works for Claude Code, Codex, OpenClaw, and Hermes). It runs the published npm package directly with npx, no clone or build required:

{
  "mcpServers": {
    "cortex": {
      "command": "npx",
      "args": ["-y", "thehive-cortex-mcp"],
      "env": {
        "CORTEX_URL": "https://cortex.example.com:9001",
        "CORTEX_API_KEY": "your-org-admin-key",
        "CORTEX_SUPERADMIN_KEY": "your-superadmin-key"
      }
    }
  }
}

The npm package is named thehive-cortex-mcp; it installs a cortex-mcp binary. Set CORTEX_SUPERADMIN_KEY only if you want the organization and user management tools.

Usage

Claude Code

claude mcp add cortex \
  --env CORTEX_URL=https://cortex.example.com:9001 \
  --env CORTEX_API_KEY=your-org-admin-key \
  --env CORTEX_SUPERADMIN_KEY=your-superadmin-key \
  -- npx -y thehive-cortex-mcp

Add --scope user to make it available from any directory instead of only the current project.

OpenClaw

openclaw mcp set cortex '{
  "command": "npx",
  "args": ["-y", "thehive-cortex-mcp"],
  "env": {
    "CORTEX_URL": "https://cortex.example.com:9001",
    "CORTEX_API_KEY": "your-org-admin-key",
    "CORTEX_SUPERADMIN_KEY": "your-superadmin-key"
  }
}'

If you are running from a source checkout instead, point command/args at the built dist/index.js:

openclaw mcp set cortex '{
  "command": "node",
  "args": ["/absolute/path/to/cortex-mcp/dist/index.js"],
  "env": {
    "CORTEX_URL": "https://cortex.example.com:9001",
    "CORTEX_API_KEY": "your-org-admin-key",
    "CORTEX_SUPERADMIN_KEY": "your-superadmin-key"
  }
}'

Then restart the OpenClaw gateway so the new server is picked up:

systemctl --user restart openclaw-gateway
openclaw mcp list   # confirm "cortex" is registered

Hermes Agent

Hermes Agent reads MCP config from ~/.hermes/config.yaml under the mcp_servers key. Add an entry:

mcp_servers:
  cortex:
    command: "npx"
    args: ["-y", "thehive-cortex-mcp"]
    env:
      CORTEX_URL: "https://cortex.example.com:9001"
      CORTEX_API_KEY: "your-org-admin-key"
      CORTEX_SUPERADMIN_KEY: "your-superadmin-key"

Then reload MCP from inside a Hermes session:

/reload-mcp

Codex CLI

Codex CLI registers MCP servers via codex mcp add:

codex mcp add cortex \
  --env CORTEX_URL=https://cortex.example.com:9001 \
  --env CORTEX_API_KEY=your-org-admin-key \
  --env CORTEX_SUPERADMIN_KEY=your-superadmin-key \
  -- npx -y thehive-cortex-mcp

Codex writes the entry to ~/.codex/config.toml under [mcp_servers.cortex]. Verify with:

codex mcp list

Standalone

export CORTEX_URL=https://cortex.example.com:9001
export CORTEX_API_KEY=your-org-admin-key
npx -y thehive-cortex-mcp     # or `npm start` from a source checkout

MCP Tools (30)

Status

Analyzer Tools

Analyzer Definition Tools

Job Tools

Responder Tools

Responder Definition Tools

Bulk Operations

Organization Management (superadmin)

User Management (superadmin)

MCP Resources (4)

MCP Prompts (4)

Configuration

Security and safety gates

This server can trigger real-world actions and submit observables to third-party services, so several capabilities are secured by default:

• Arbitrary file reads are blocked. cortex_run_analyzer_file only reads files inside CORTEX_FILE_BASE_DIR (realpath-confined to defeat symlink/.. escapes). With no base dir configured, path-based reads are refused; use fileBase64 to submit content explicitly.
• Writes are opt-in. Set CORTEX_MODE=write before submitting analysis or changing configuration. POST job searches remain available in read-only mode.
• Responders are gated. cortex_run_responder requires both CORTEX_ALLOW_DESTRUCTIVE=1 in the environment and confirm=true in the call.
• Changes require confirmation where indicated. User creation with roles, key rotation, organization locking, responder removal, job deletion, analyzer removal and cleanup with dryRun=false require confirm=true. Deletion and responder execution also require CORTEX_ALLOW_DESTRUCTIVE=1.
• Bulk analysis is conservative. cortex_analyze_observable does not fan out to every analyzer by default. Pass an explicit analyzers allowlist, or set fanOut=true to run all applicable analyzers (capped by CORTEX_MAX_FANOUT).
• API transport requires HTTPS. HTTP is allowed without an override only for literal dotted-decimal addresses in 127.0.0.0/8 and IPv6 ::1 (including expanded notation). DNS names, including localhost and names ending in .localhost, IPv4-mapped IPv6, encoded addresses and legacy IPv4 forms require the explicit insecure HTTP override. No DNS lookup determines this exception. Prefer HTTPS with a trusted private CA through NODE_EXTRA_CA_CERTS. The override does not disable certificate verification or permit URL userinfo.
• SSL verification is scoped. Disabling CORTEX_VERIFY_SSL relaxes TLS only for Cortex connections, never for the whole Node process.

Result and input limits

Responses are limited to 4 MiB while streaming. JSON request bodies are limited to 1 MiB, incoming MCP message lines to 16 MiB before parsing, and file uploads to 10 MiB. Path-based files must be regular files within the configured base directory. Text arguments are limited to 65,536 characters, with a separate encoded-file limit. Tool and resource data is JSON-quoted inside escaped untrusted_cortex_data blocks, limited to 64 KiB with a truncation marker. Resources use text/plain for that envelope. Upstream error bodies and status text are omitted. Credential-bearing configuration and serialized parameters are omitted. Common credential fields are redacted from structured results.

Credential delivery and worker networking

The Proxmox installer reads keys silently, sends credentials through stdin, and stores them with owner-only permissions for systemd LoadCredential. The server reads cortex-api-key and cortex-url from CREDENTIALS_DIRECTORY when their respective environment variables are absent. Environment configuration still works.

The installed service runs as the dedicated cortex-mcp system account with no login shell. Application files under /opt/cortex-mcp remain root-owned and read-only to that account. Source credentials under /etc/cortex-mcp remain root-only. Systemd supplies private credential copies and owns the service's writable state and runtime directories (/var/lib/cortex-mcp and /run/cortex-mcp, mode 0700). The unit uses a read-only filesystem sandbox, a private temporary directory, no capabilities and NoNewPrivileges. Keep private CA certificates readable by the account when configuring NODE_EXTRA_CA_CERTS in a unit override.

For embedded use, pass an operator-approved CredentialStore adapter as the second CortexClient constructor argument. Its put(userId, key) method must persist the key without logging it and return a non-secret reference. Rotation returns only that reference and a SHA-256 fingerprint. A failed store write is reported without key material. If it fails after renewal, recover through Cortex administration.

Network-target validation covers IP, domain, FQDN and HTTP/HTTPS URL submissions. URL credentials and unsupported schemes are refused. DNS checks have a five-second budget and reject any private or special-purpose answer unless explicitly permitted. Configure the Cortex workers' network policy to enforce the same destination rules on every connection, including redirects and later DNS resolutions. Submission-time checks cannot enforce worker-side redirects or DNS changes. Cortex API redirects are refused by the client.

Examples

Set up analyzers from scratch

1. Use cortex_list_analyzer_definitions with freeOnly=true to find analyzers
   that need no API keys.
2. Use cortex_enable_analyzer to enable "Abuse_Finder_3_0" with empty config.
3. Use cortex_analyze_observable with data "8.8.8.8" and fanOut=true to run
   all applicable analyzers (or pass analyzers ["Abuse_Finder"] to scope it).

Auto-detect observable type

Use cortex_analyze_observable with data "185.220.101.42" and fanOut=true
(no dataType needed - auto-detects as IP). Or pass an `analyzers` allowlist
to limit which analyzers run.

Clean up old failed jobs

Use cortex_cleanup_jobs with status "Failure", dryRun true to preview,
then dryRun false and confirm true to delete (requires write mode and destructive permission).

Analyze a file

Set CORTEX_FILE_BASE_DIR=/srv/cortex-uploads in the server environment, then:
Use cortex_run_analyzer_file with analyzerId "Yara_3_0",
filePath "/srv/cortex-uploads/suspicious.exe" to scan with YARA rules.
(Paths outside CORTEX_FILE_BASE_DIR are refused; alternatively pass fileBase64.)

Manage API keys

Key retrieval is unavailable. Rotation requires write mode, confirm=true and an
operator-configured credential store adapter. The stock stdio server has no adapter
and refuses rotation before contacting Cortex. Manage keys through Cortex administration
until a store adapter is configured.

Triage a security alert

Use the triage-alert prompt with alertDescription "Suspicious outbound traffic
detected" and observables "185.220.101.42, evil.example.com, 44d88612fea8a8f36de82e1278abb02f"

Supported Data Types

Deployment

Proxmox LXC

Before running the installer, review a local copy of the script and set these immutable inputs. There are no defaults:

• CORTEX_SOURCE_COMMIT: the full 40-character hexadecimal commit reviewed for installation from https://github.com/lidless-labs/cortex-mcp.git.
• NODE_VERSION: an exact Node 22 version at least 22.12.0, without a v prefix.
• NODE_ARCH: x64 or arm64, matching the guest architecture.
• NODE_SHA256: the 64-character SHA-256 of that version's node-v<VERSION>-linux-<ARCH>.tar.xz archive.

Obtain the archive checksum from the official release's signed SHASUMS256.txt.asc. Verify its signature with a separately trusted Node release keyring using the official binary verification instructions before supplying the checksum. Do not copy a checksum from the same unverified download as the archive.

The installer rejects missing or malformed pins before any host command or LXC creation. It downloads only from https://nodejs.org/dist/v<VERSION>/, checks the archive against the supplied checksum before extraction, then fetches the reviewed source commit from the canonical repository and checks its exact hash before npm ci --ignore-scripts or the build. Node is installed in /usr/local, including the service executable /usr/local/bin/node. Ubuntu template and apt prerequisites still follow their upstream repositories. This does not freeze the whole OS or provide a reproducible apt snapshot. An interrupted install leaves its guest for operator inspection. A retry refuses an existing container ID.

Release publishing

The tag workflow uses npm trusted publishing and provenance. Configure the package's trusted publisher for lidless-labs/cortex-mcp and .github/workflows/ci.yml on npm before the next release. The workflow uses Node 24 and grants OIDC permission only to the publish job. It does not use an npm token. Changing this workflow does not publish a release or configure the npm account.

Why not something else?

• The Cortex web UI is built for one analyst clicking through jobs by hand. cortex-mcp puts the same engine behind an AI client, so analysis, taxonomy aggregation, and artifact extraction happen conversationally instead of through a dozen page loads.
• A raw REST wrapper or generic HTTP MCP bridge gives a model an untyped endpoint and no domain knowledge. cortex-mcp models analyzers, responders, jobs, definitions, organizations, and users as 31 typed tools with auto data-type detection, capped fan-out, and built-in safety gates, so the agent works in Cortex's vocabulary rather than reconstructing the API from scratch.
• Wiring Cortex into a SOAR runbook or n8n flow is great for fixed, pre-authored pipelines. This server is for the open-ended path: ad hoc enrichment, investigation, and triage where the analyst (or the agent) decides the next step as results come in.
• thehive-mcp (the companion server) drives case and alert management in TheHive. cortex-mcp is the analysis-and-response layer; the two are complementary, not substitutes.

What cortex-mcp is not

• It is not a Cortex replacement or a reimplementation of analyzers. It calls a Cortex instance you already run; Cortex still does the analysis.
• It is not a SIEM, a case manager, or a TheHive client. Case and alert workflows belong to TheHive (see thehive-mcp).
• It is not an autonomous responder. Destructive actions (responders, job deletion, file reads by path) are off or confirmation-gated by default and never fire silently.
• It is not a hosted service. It runs locally as a stdio MCP server next to your client; nothing is sent anywhere except the Cortex instance you configure.

Testing

npm test              # Unit tests
npm run test:watch    # Watch mode
npm run lint          # Type check

# Integration tests (requires live Cortex instance)
CORTEX_URL=https://cortex.example.com:9001 \
CORTEX_API_KEY=your-key \
CORTEX_SUPERADMIN_KEY=your-superadmin-key \
npx vitest run tests/integration.test.ts

Live lifecycle tests also require CORTEX_TEST_MUTATIONS=true, CORTEX_MODE=write and CORTEX_ALLOW_DESTRUCTIVE=1. Use a disposable Cortex instance for these tests. They remove the analyzers and responders they create and report cleanup failures.

Project Structure

cortex-mcp/
  src/
    index.ts                  # MCP server entry point
    config.ts                 # Environment config + validation
    client.ts                 # Cortex REST API client (full surface)
    types.ts                  # Cortex API type definitions
    resources.ts              # MCP resources (4)
    prompts.ts                # MCP prompts (4)
    tools/
      analyzers.ts            # Analyzer tools (list, get, run, run-by-name)
      analyzer-definitions.ts # Definition browsing, enable, disable
      jobs.ts                 # Job management + cleanup
      responders.ts           # Responder tools (list, run)
      responder-definitions.ts # Definition browsing, enable, disable
      bulk.ts                 # Bulk analysis with auto-detect
      status.ts               # Health/version check
      organizations.ts        # Org CRUD (superadmin)
      users.ts                # User CRUD + key management (superadmin)
  tests/
    client.test.ts            # API client unit tests
    tools.test.ts             # Tool handler unit tests
    integration.test.ts       # Live instance integration tests
  scripts/
    proxmox_install.sh        # Proxmox LXC deployment script

Contributing

Issues and pull requests are welcome. See CONTRIBUTING.md for the contribution path and SECURITY.md for reporting vulnerabilities. By participating you agree to the Code of Conduct.

Container root access

The installer creates no guest root password and locks password login after starting the container. Manage the guest from its Proxmox host with pct enter. Service API credentials continue to use protected stdin delivery and root-only storage.

License

MIT. See LICENSE.

Part of Lidless Labs · the eye does not close

_{Security / SOC: soc-stack · wazuh-mcp · misp-mcp · suricata-mcp · thehive-mcp · mitre-mcp · zeek-mcp · hotwash}

_{All tools · Lidless Labs on GitHub}