SentinelOne logo
Live

SentinelOne MCP server and AI skills

Every endpoint SentinelOne protects — health, threats, policy — answerable in plain English.

SentinelOne is the endpoint protection on your clients' machines, and the console knows things your RMM cannot: whether the agent is actually online and current, which threats are open and for how long, and which sites are on a detect-only policy. Connect your console and a session can read all of it across every site — agents, threats and their timelines, sites and licences, groups, policies and exclusions, application inventory and vulnerability counts. Read-only today: no credential we issue carries a write scope, and no connected session can reach a write tool, so nothing here isolates a machine, kills a process, changes a policy or edits an exclusion. Switchboard connects through an MSPStuff-hosted MCP server, so there is nothing to install and no API token to hand your techs.

  • 24 skills
  • 6 playbooks
  • Read-only by default
  • Every answer traced
See the shape of an answer

A question in plain English. A result you can work with.

This illustrative session shows one of the documented questions below. Your results come from your connected environment and the access your admin grants.

SentinelOne · illustrative session
“Which sites have agents that have not checked in for 30 days?”
These sites have agents that stopped checking in more than 30 days ago.
Sites with agents silent for 30 days
SiteAgents past 30 days
Acme6 of 148
Harbor3 of 92
Northstar2 of 61
Example source: SentinelOne · scoped read
Sample data. Illustrative results, not a customer report.

The story

What changes once it’s connected.

Hosted and external AI connections provide read-only access. Guidance, shared playbooks and automations we release for everyone are included with Dedicated. Bespoke automation builds are scoped separately. Any change a platform could make is held back — refused by our server before anything reaches the platform — until a person switches it on, and nothing is switched on today. Vendor consent does not switch any change on.

"Installed" is not "protected". Your RMM can tell you the agent is on the box; only SentinelOne can tell you it is online, up to date, scanned, and on a policy that blocks rather than just detects. Connected, those get asked once and answered across the whole book of clients, ranked by site, instead of one console filter at a time.

The coverage follows how SentinelOne is really structured. Sites — the clients — with their licences and state, groups beneath them with the policy each inherits or overrides, and agents with everything the console holds about their health. Threats as incidents with a mitigation status, an analyst verdict and a timeline you can read end to end. Exclusions and blocklists with the scope they apply at, because a broad path exclusion at account scope is the kind of thing nobody remembers adding.

It also closes the loop with your RMM. When Automate or NinjaOne is connected too, a session can name the machines the RMM manages that carry no SentinelOne agent, and the agents SentinelOne has that the RMM has no record of — both halves of the onboarding gap. Nothing here acts on a threat or a machine; to isolate, quarantine or change a policy, a human uses the SentinelOne console. Any change the integration could make in SentinelOne is held back — refused by our server before anything reaches SentinelOne — until a person switches it on, and nothing is switched on today.

Things you can ask

Ask it like you’d ask a teammate.

>Which sites have agents that have not checked in for 30 days?
>What threats are still unresolved, and which clients are they on?
>Which machines in Automate have no SentinelOne agent?
>Are any sites running detect-only instead of protect?
>How many agents are behind the console's latest version, by site?
>What happened on that threat — walk me through its timeline.
>Which exclusions apply at account scope, and who added them?

Skills and playbooks

Every skill, in the open.

24skills6playbooksLast vetted Sep 24, 2026

Read-only today, across every site the service user can see: agents and their health, threats and their timelines, sites and licences, groups, policies and exclusions, application inventory and vulnerability counts. No credential we issue carries a write scope, and no connected session can reach a write tool — so there is no path from a session to isolating or disconnecting a machine, killing or quarantining anything, changing a policy, editing an exclusion, or running a script. The credential is a service user you create in your own console with the Viewer role, so everything a session sees is exactly what that user can see, and every read is attributed to it, in your console and in ours. Any change the integration could make in SentinelOne is held back — refused by our server before anything reaches SentinelOne — until a person switches it on, and nothing is switched on today.

What it can do

  • Sites & licensingThe shape of the console: the account, its sites (one per client, usually) with state and licence counts, and the groups beneath each site — the resolver for which client a question is about.4skills

    Reading only — no site, group or licence is created or changed from here.

    • The account the token sees, with its licence and agent totals — the top of the tree before you narrow to a client.
    • Every client site with its state and licences, which is how a customer name becomes the site a session then asks about.
    • One site opened in full — licences, expiry and settings behind the name, when the roster row is not enough.
    • The groups under each site with their agent totals and whether each inherits its policy or overrides it — where a machine that behaves differently usually turns out to be; what rule a dynamic group uses to decide its membership stays in the console and is not readable here.
  • Agents & healthThe endpoint roster and its health, from a one-call count of any slice to one machine's full record — online, current, scanned, decommissioned, last seen.5skills

    Reading only — no agent is isolated, disconnected, scanned or sent anything. Health flags are triage filters, not actions.

    • How many agents match a slice — offline, out of date, never scanned, silent for a month — in one call, without paging a single machine.
    • Page through agents by site, group, state or last check-in, so a ranking of who is worst comes from the console's own filters — with two things they will not do: there is no filter for a machine waiting on a reboot, so a fleet-wide count of those only exists if the whole roster is walked, and a state the console does not recognise matches nothing and says so rather than reading as an empty fleet.
    • One machine opened in full — the whole record behind an offline flag, not the summary row.
    • What is installed on specific machines, for the moment an advisory lands and the question is which endpoints carry it.
    • The console's own release and the version it advertises as the latest agent — worth knowing, and not a yardstick: on a live fleet the console advertised an older release than the machines were running, so whether an agent is out of date is read from the agent's own current-version flag, never from this number.
  • Threats & response historyOpen incidents first, then the history: what was detected, what the engine did about it, what an analyst decided, and the timeline and notes behind one threat.5skills

    Reading only — no threat is mitigated, resolved, marked benign or annotated from here.

    • Unresolved threats across every site, with classification, mitigation status and verdict — the open list a QBR starts from; the console spells its classifications exactly, so a filter that matches nothing says so rather than reporting a clean fleet.
    • One threat opened in full — indicators, detection detail and mitigation steps — when the summary row leaves a question open; the raw process telemetry behind a detection belongs to deep visibility, a separate product that keeps thirty days, and is not on this surface.
    • What happened to one threat in order — detection, mitigation, verdicts — so the story can be told without opening the console, read a page at a time because the console refuses a page smaller than ten events and caps one at two hundred.
    • The analyst notes on a threat, for the context a mitigation status alone leaves out.
    • Custom-rule alerts with the process and rule behind each, for the detections that are not classic threats.
  • Policies & exclusionsWhat protection is actually configured: the policy at account, site and group scope, the exclusions and the blocklist with the scope each applies at, and the console's own activity log.7skills

    Reading only — no policy, exclusion or blocklist entry is created, edited or removed from here.

    • The account-wide policy — mitigation modes and engines — that every site inherits unless it overrides.
    • One site's effective policy, which is where detect-only shows up when a client was onboarded cautiously and never moved on.
    • One group's effective policy, for the group that overrides its site and the machine that behaves differently because of it.
    • Every exclusion the account layer carries, with its type and the scope it applies at — the broad path exclusion nobody remembers adding; a site's own exclusions are read site by site, and the group layer is not swept, so no single read is every exclusion in the fleet.
    • The blocklist and where each entry applies, for the question of what is being stopped on purpose.
    • The console's own audit log — who changed what and when — narrowed to a week, a site or an activity type.
    • The catalogue of activity types, so a question about policy changes or agent installs can be asked by name — summarised by default, because a console carries hundreds of them and the full catalogue will not fit in one answer; asked for in full it comes a page at a time.
  • Applications & vulnerabilitiesThe application inventory SentinelOne sees on every agent and the vulnerability counts it computes from it, read as counts and narrow slices rather than as a firehose.2skills

    Reading only — nothing is patched, uninstalled or scanned; a vulnerability count is a fact, not an instruction.

    • Which machines carry an application, by name — the fastest answer when a vendor advisory names a product.
    • Application vulnerabilities by severity and site, counted at the source rather than walked — always with a severity or a site named, because an unfiltered count makes the console walk millions of rows and can take the best part of a minute, and because a filter that matches nothing says so rather than reporting a clean fleet.
  • Coverage & reconciliationWhere protection is missing: network devices the console discovered that carry no agent, and — with an RMM connected — the machines the RMM manages that SentinelOne has no record of.1skills

    Reading only — no device is enrolled, deployed to or removed from here.

    • Devices seen on the network that no agent covers, on consoles that license network discovery — where a licence is not the same as a result: on the fleet we vetted, discovery is licensed and the console has seen nothing, so the honest answer there is none rather than a gap.

How it has been taught to work it

Playbooks are the vetted techniques a session loads for this platform — what to check, in what order, and what a number means before it is reported.

  • Agent health — offline, out of date, silent, never scannedWhich sites have agents offline, out of date, silent or never scanned; is this machine protected; agent version questions.
    Vetted Sep 24, 2026
  • Coverage — RMM machines with no SentinelOne agentWhich machines in Automate or NinjaOne have no SentinelOne agent; agents the RMM has never seen; sites with no agents; licence overuse; onboarding gaps.
    Vetted Sep 24, 2026
  • Open threats — unresolved, unmitigated, agingWhat threats are open, unmitigated or aging, on which clients; what happened on one threat; weekly or quarterly threat summary for a QBR.
    Vetted Sep 24, 2026
  • Policy drift — detect-only sites, exclusions and the blocklistWhich sites are detect-only; what a site's or group's policy is; exclusions and blocklist by scope; who changed a policy and when.
    Vetted Sep 24, 2026
  • Querying SentinelOne — the cache, cursor paging and the filters it honoursAlways when SentinelOne is attached. Which questions the cache answers in one call, how cursor paging works and where it caps, which agent and threat filters the tenant honours, what a Viewer is refused, and the traps found live.
    Vetted Sep 24, 2026Always on
  • SentinelOne — what the platform ISAlways when SentinelOne is attached. Its entity model — account, sites, groups, agents, threats vs incidents vs alerts, policies and inheritance, exclusions — the definitions the mirror computes, and why installed in the RMM is not protected here.
    Vetted Sep 24, 2026Always on

Plus 18 more every session carries regardless of platform — how to shape an investigation, keep every number traceable, and say plainly when something is broken.

Connecting

Wire it in, in an afternoon.

How to connect

  1. 01In your SentinelOne console, create a service user (Settings → Users → Service Users) with the Viewer role on your account and an expiry date.
  2. 02Copy the API token when it is shown — it is shown once — and note your console URL, the sentinelone.net address you sign in at.
  3. 03Paste the console URL and the token on the connect page — it validates live against SentinelOne before saving.

Common questions

Asked before you had to ask.

Does SentinelOne have an API or MCP server?

SentinelOne provides RESTful and GraphQL APIs and publishes the open-source Purple AI MCP server, which is read-only and which you run yourself. MSPStuff hosts a SentinelOne MCP server for you: across every site your service user can see, read-only today, with every answer traced to the real tool calls behind it.

Is there an AI skill for SentinelOne?

Yes — MSPStuff ships AI skills for SentinelOne, hosted in Switchboard. You connect your console once; sessions then answer questions about it in plain English. There is nothing to install to use Switchboard and no API token to hand your techs.

Is there a SentinelOne MCP server?

Yes. The SentinelOne integration is an MCP server that MSPStuff hosts and operates for you — Switchboard sessions call it with a credential you control, and every answer is traced to the real tool calls behind it.

Can it tell me which machines have no SentinelOne agent?

Yes, two ways. The console's own network discovery lists devices it sees with no agent, and with an RMM such as ConnectWise Automate or NinjaOne connected too, a session can name the machines the RMM manages that SentinelOne has no record of.

Can it change anything in SentinelOne?

No. Read-only today: no credential we issue carries a write scope, and no connected session can reach a write tool, so nothing here isolates a machine, kills a process, changes a policy or edits an exclusion. Any change the integration could make in SentinelOne is held back — refused by our server before anything reaches SentinelOne — until a person switches it on, and nothing is switched on today. See the access section above for the exact guarantee.

Does this use the SentinelOne API?

Yes. The SentinelOne integration is an MSPStuff-hosted MCP server built on the SentinelOne Management API. Read-only today: no credential we issue carries a write scope, and no connected session can reach a write tool. You connect once with a service-user token you control; nothing is installed on your side.

Is there a SentinelOne connector for my AI?

Yes. Supported clients such as ChatGPT (OpenAI), Claude, Copilot and Cursor use the same MSPStuff MCP endpoint. Your admin controls access per person and data domain, and every platform call uses read-only permissions.

Get AI connected to your stack.

Apply for the beta: 14 days free on the Hosted plan against your own environment, read-only, every answer traced. Accepted MSPs get half off seats and client tenants for the first year. Dedicated infrastructure is priced separately.

MSPStuff