wrg-mcp-server

💡 Found this useful? ⭐ Star the repo (helps others find it) and subscribe to weekly detection-engineering writeups at Detection Frontier.

Give your AI agent direct access to a 60+ tool security & threat-intel stack via MCP — secret scanning, sigma rule generation, ransomware lookup, OSINT, deep research, and more.

An MCP (Model Context Protocol) bridge that exposes the WinstonRedGuard AI security platform to Claude Code, Claude Desktop, Cursor, Codex, and any MCP-compatible AI agent.

Why this exists

Modern AI coding agents (Claude, Cursor, Codex) excel at code generation but cannot directly:

• Scan a repo for leaked secrets, credentials, or PII patterns

• Look up ransomware groups, victims, and infrastructure by name

• Generate Sigma detection rules from observed incident data

• Run OSINT username probes across 3000+ sites

• Query a curated threat-intel actor corpus with MITRE ATT&CK mappings

This MCP server fills that gap. Drop it into your MCP client config and your agent gains 60+ tools spanning code security, threat intelligence, OSINT, research, and signal analysis — all without leaving the agent's context.

Use cases

• Secure-by-default coding workflows — agent runs devguard_scan before every commit, catches API keys + PII patterns before they ship

• Threat-intel queries during incident response — "What's the latest LockBit infrastructure?" → agent calls ransomware_lookup + darkweb_brand_watch directly

• Detection engineering with AI assistance — ai_fingerprint_sigma_emit converts observed AI-generated code patterns into Sigma YAML rules

• OSINT investigations — maigret_search username probe across 3000+ sites; results flow back into the agent's reasoning context

• Research automation — research_motor HTTP API gives the agent persistent, queryable research jobs across domains

Quick start

pip install wrg-mcp-server                 # core: 40+ local tools
pip install "wrg-mcp-server[remote]"       # adds httpx for site_* / pulseboard_* tools

Add to your Claude Code / Claude Desktop config:

{
  "mcpServers": {
    "wrg": {
      "command": "wrg-mcp-server",
      "args": ["--transport", "stdio"],
      "env": {
        "WRG_MCP_ALLOW_MUTATIONS": "0"
      }
    }
  }
}

Restart your client. The agent now has access to mcp__wrg__* tools.

How it compares

When to reach for wrg-mcp-server

• You're a security engineer, detection engineer, or threat-intel analyst working with AI agents

• You want curated threat-intel + OSINT in your agent without manual tool-juggling

• You have (or can install) the WinstonRedGuard monorepo for the full feature set

Where wrg-mcp-server loses today (honest delta)

• Setup friction higher than generic MCP servers — full feature set requires the WinstonRedGuard monorepo (private). Standalone install (no monorepo) gives ~40% of tools (OSINT + research + trading + polymarket all work without monorepo)

• Windows-first — primary dev environment is Windows 11 + WSL2; macOS/Linux supported but less battle-tested

• Documentation density — 60+ tools is a lot; in-depth per-tool docs live in source comments rather than separate pages

• Newer than alternatives — github-mcp-server, Filesystem, Playwright MCP all have larger communities and more battle-testing

Transports

wrg-mcp-server --transport stdio              # Claude Desktop / Claude Code (recommended)
wrg-mcp-server --transport streamable-http    # HTTP for remote clients
wrg-mcp-server --transport sse                # legacy HTTP (SSE)

Flags: --host 0.0.0.0 · --port 8080 · --mcp-path /mcp

Install (full options)

pip install wrg-mcp-server                 # core: MCP + local tools only
pip install "wrg-mcp-server[remote]"       # adds httpx for site_* / pulseboard_* tools
pip install "wrg-mcp-server[dev]"          # pytest + pytest-asyncio

From source (standalone repo):

git clone https://github.com/WRG-11/wrg-mcp-server.git
cd wrg-mcp-server
pip install -e ".[dev]"

Note: [threat-intel] extras were removed in v1.0.4 (PyPI rejects direct file:// deps). Sister wrg_threat_intel + ransom_radar stay in the WRG monorepo for now; will re-add this extras group once they publish to PyPI.

Tool surface

60+ tools organised across 8 categories. Detailed tables below — expand a section to view.

Core monorepo introspection

Research

Silo-app expansion (6 apps × 2 tools)

AI fingerprint (wrg_ai_fingerprint)

DevGuard (wrg_devguard)

Security suite (wrg_security_suite) — security_suite_run is mutation-gated

Rule lab (rule_lab)

Data janitor (data_janitor) — data_janitor_sweep mutation-gated when dry_run=False

Notifier (wrg_notifier3) — notifier_send is mutation-gated

INFO_OPS extension

Example:

info_ops_detect()  # all INFO_OPS actors + Sigma + incidents
info_ops_detect(actor_id="russia_nexus_info_ops")  # specific actor
info_ops_detect(mitre_technique="T1656")  # reverse lookup

OSINT

Threat-intel (opt-in via [threat-intel] extra)

Ransom-radar

Remote tools return {"ok": false, "error": "httpx not installed — remote tools unavailable"} when [remote] extra is missing.

Environment

Repo discovery

Mutation gate (default: off)

State-changing tools (memory_set, pipeline_run, security_suite_run, data_janitor_sweep non-dry, notifier_send) refuse to execute unless:

WRG_MCP_ALLOW_MUTATIONS=1

This prevents an MCP client from silently writing memory or launching pipelines on a read-only connection.

Remote service config

Per service (SITE / PULSEBOARD), prefix with WRG_<SERVICE>_:

research_motor HTTP API

Start the research_motor API separately, then point the MCP server at it:

cd apps/research_motor
pip install -e ".[api]"
set RESEARCH_MOTOR_API_KEY=replace-me
research-motor serve --host 127.0.0.1 --port 8080

Configure the MCP server environment:

set WRG_RM_API_BASE_URL=http://127.0.0.1:8080
set WRG_RM_API_KEY=replace-me

Architecture

FastMCP server
├── server.py            — tool registration, remote HTTP dispatch
├── config.py            — ServiceConfig / AppConfig from env (frozen dataclasses)
├── http_utils.py        — URL builder, response parser
├── local_tools.py       — subprocess wrappers for WRG CLIs (~20 tools)
└── cli.py               — argparse entry point

Local tools use subprocess.run with stdin=DEVNULL (not asyncio subprocess) — avoids a Windows pipe-blocking deadlock under anyio. Tool dispatch is wrapped in anyio.to_thread.run_sync so the MCP event loop stays responsive.

Tests

pytest -q

Sister WRG-11 packages

Part of the WRG-11 PyPI portfolio:

• instinct-mcp — Self-learning memory for AI coding agents

• wrg-devguard — Developer-first AI safety: prompt-policy lint + secret scanning + log scanning with PII detection

• wrg-rule-lab — Local-first deterministic rule evaluation engine (zero-dep, stdlib-only)

• ai-security-toolkit — Offensive + defensive AI/LLM security tools, labs, CTF writeups, research

Built by WRG-11.

Status

Production — covers every active WRG app, drives the mcp__wrg__* tools visible in connected Claude sessions.

License

MIT. See LICENSE.