wrg-mcp-server 1.0.2

WRG MCP server - exposes WinstonRedGuard tools to Claude and AI agents

wrg_mcp_server

MCP (Model Context Protocol) server exposing the WinstonRedGuard monorepo to Claude and other MCP-compatible AI agents. Built on FastMCP — registers tools from every active WRG app so an agent can inspect the repo, run pipelines, query memory, and call remote services without shelling out.

Published:

• PyPI: wrg-mcp-server (v1.0.1, 2026-04-25)
• MCP Registry: io.github.yakuphanycl/wrg-mcp-server (status: active, 2026-04-25)

Transports

wrg-mcp-server --transport stdio              # Claude Desktop / Claude Code
wrg-mcp-server --transport streamable-http    # default HTTP (recommended)
wrg-mcp-server --transport sse                # legacy HTTP

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

Install

cd apps/wrg_mcp_server
pip install -e .             # core: MCP + local tools only
pip install -e ".[remote]"   # adds httpx for site_* / pulseboard_* tools
pip install -e ".[dev]"      # pytest + pytest-asyncio

Tools exposed

Local (subprocess-backed, always available)

Tool	What it does
connector_status	Report which remote services are configured
app_list, app_info	Query app_registry/data/registry.json
governance_run	Execute governance_check across one or all apps
release_check	Run the tools/release_check.ps1 gate
pipeline_list, pipeline_show, pipeline_run	wrg_pipeline DAG operations
pulse_check	Invoke wrg-pulse check
memory_get, memory_set, memory_list, memory_search	wrg_memory key-value access
research_history, research_report, research_scan, research_watch, research_scan_summary	research_motor runs and artifacts
research_motor_healthz, research_motor_scan_create, research_motor_scan_get	research_motor HTTP API v1 over localhost
vault_audit	wrg_vault audit ledger inspection
scheduler_task_list, scheduler_tick_dry_run	wrg_scheduler inspection

Remote (HTTP, opt-in via env)

Tool	Upstream
site_health, site_get, site_post	Company site API (WRG_SITE_BASE_URL)
pulseboard_health, pulseboard_list_repos, pulseboard_add_repo, pulseboard_delete_repo, pulseboard_get_pulse	pulseboard dashboard (WRG_PULSEBOARD_BASE_URL)

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

research_motor HTTP API tools

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

Example MCP tool payloads:

{"target": "example.com", "mode": "domain"}

{"scan_id": "sample-scan-abc123"}

Environment

Repo discovery

Variable	Default	Purpose
WRG_REPO_ROOT	auto-detect (walk up until apps/ + CLAUDE.md)	Required when installed from wheel outside the monorepo

Mutation gate (default: off)

State-changing tools (memory_set, pipeline_run) 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>_:

Variable	Default	Purpose
*_BASE_URL	—	Enables the service (unset = service disabled)
*_TOKEN	—	Bearer token for Authorization header
*_AUTH_HEADER	Authorization	Override header name
*_AUTH_SCHEME	Bearer	Override token scheme
*_SESSION_COOKIE	—	Optional Cookie header
*_EXTRA_HEADERS	—	JSON object of extra headers
*_TIMEOUT_SECONDS	WRG_HTTP_TIMEOUT_SECONDS (20.0)	Per-request timeout
*_VERIFY_TLS	WRG_HTTP_VERIFY_TLS (true)	TLS verification

Claude Code / Claude Desktop integration

Add to your MCP client config:

{
  "mcpServers": {
    "wrg": {
      "command": "wrg-mcp-server",
      "args": ["--transport", "stdio"],
      "env": {
        "WRG_REPO_ROOT": "D:\\dev\\WinstonRedGuard",
        "WRG_MCP_ALLOW_MUTATIONS": "0"
      }
    }
  }
}

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

Status

Production — 1045 lines, covers every active WRG app, drives the mcp__wrg__* tools visible in connected Claude sessions.