MCP Server — AI-Driven Automation and Repair
RMMmax ships a Model Context Protocol (MCP) server exposing roughly 79 canonical tool functions over a secure JSON-RPC endpoint, letting an AI assistant — Claude, or any MCP-compatible agent — inventory, diagnose, script, and remediate a customer’s entire managed fleet, with the same tenant isolation and entitlement rules enforced everywhere else in the platform.
Authentication & Security Model
- OAuth2
client_credentialsflow — operators create scoped OAuth clients from Environment → MCP Server, receiving a client ID/secret pair once. - Three explicit scopes — read (inventory, status, summaries, docs), write (enable/disable clients, agents, tool configs; create scripts), execute (dispatch commands/scripts, run scans/updates) — fail-closed by default. An empty scope grants nothing.
- Every tool call is re-scoped to the authenticated team; cross-tenant IDs are returned as “not found,” never dispatched.
- Rate-limited to 120 requests/minute per credential; browser-origin requests are restricted to allowed AppBox origins.
- Responses use explicit field allowlists — credentials, OAuth secrets, agent tokens, and encrypted fields are never returned through MCP, even to a fully-scoped client.
- MCP access is plan-gated (an add-on tool slug), and every individual tool call is additionally checked against the team’s plan entitlements — a team without the BitLocker entitlement cannot call BitLocker MCP functions even with a full-scope token.
Function Categories
| Category | Representative Functions |
|---|---|
| Inventory & docs | List endpoints/clients, inventory summary, CVE summary, read/search platform documentation |
| Alerts | List, acknowledge, and create alerts |
| Scripts & executions | List/create/run scripts and commands, check execution status, enable/disable clients, agents, and tool configurations |
| Software management | Get Chocolatey/Homebrew/Linux package data, install/update packages, reboot agents |
| Health & Defender | Health summaries/detail, virus quick/full scans, Defender status |
| BitLocker | Scan, lock/unlock/suspend/resume volumes, client/agent data |
| BleachBit | Scan and clean actions, client/agent data |
| SurfLog | Usage data, hourly/daily browser usage |
| Update Manager | Scan for updates, get update data, trigger OS/agent updates |
| Warranty | Scan for warranty, get client/agent warranty data, list providers, get coverage, force refresh |
| Printers, Hardware Inventory, Recon | Read fleet hardware and printer detail, trigger and check network recon scans |
The Detect-Decide-Repair Loop
Because these functions are fully composable, an AI agent connected via MCP can, in one conversational session: identify every Windows agent with an overdue critical CVE, confirm which of those already has BitLocker enabled, dispatch a targeted patch run, verify the patch succeeded, and generate a client PDF report — the same detect-decide-repair loop a senior technician would run manually, done end to end through natural language.