Enable the MCP server

Everwas ships an MCP server so an AI assistant can query and, with explicit confirmation, act on your fleet. It is off by default and requires a scoped API key.

Why this exists

An RMM knows things that are tedious to ask a database directly:

Which machines had an outdated OpenSSL last Monday, and did we know that at the time?

That question needs both time axes of the bitemporal fact store, which is exactly what get_device_facts exposes. See query device history for how to phrase these.

Enabling it

.env

EVERWAS_MCP_ENABLED=true

docker compose --profile mcp up -d everwas-mcp

The MCP endpoint gets its own hostname (EVERWAS_MCP_DOMAIN, defaulting to mcp-<your-domain>) rather than a path on the app origin. The reason is the auth model: browsers talk to the app with a session cookie, MCP clients present a bearer API key, and keeping the origins separate keeps the cookie away from the API-key surface entirely.

Mint a key with only the scopes you mean

make api-key NAME=claude SCOPES=devices:read,alerts:read,patches:read
# prints once: ewpk_<id>_<secret>

Available scopes: devices:read, alerts:read, alerts:write, scripts:run, patches:read, patches:write.

Scope enforcement is server-side. A read-only key cannot run scripts, no matter what the assistant is asked or how persuasively it asks.

Connecting

claude mcp add --transport http everwas https://mcp-rmm.example.com/mcp \
  --header "Authorization: Bearer ewpk_..."

In dev the server listens on http://127.0.0.1:28001/mcp.

The safety model

Three properties, and they compose:

1. Scopes cap capability. The key decides what is possible; the assistant’s instructions decide what is attempted.
2. Mutations are two-step. Every mutating tool takes confirm: bool = False and, without it, returns a dry-run preview of exactly what it would do. Nothing changes until the assistant is told to confirm, which means you can read the plan first.
3. Everything is audit-logged. Every call, successful or refused, writes an audit row naming the key as the actor. Anything an assistant did is visible in the Audit view alongside human actions.

The full tool list with scopes and semantics is in the MCP tool reference.