defender-xdr
Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type
@followed by the MCP server name and your instructions, e.g., "@defender-xdr Show me the latest high-severity incidents from the last 24 hours."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
ZeroSOC MCP Servers
Model Context Protocol (MCP) servers, and the typed tool clients they wrap, for security operations.
MCP servers are the integration layer that lets AI agents query security platforms directly — the source of truth — instead of copying telemetry into yet another data lake. This repository is the ZeroSOC Framework community's home for MCP servers covering major security technologies that lack an official or complete one.
Operating model
For each security technology we pick exactly one engagement mode:
Mode | When | What we do |
Adopt | An official vendor (or healthy third-party) MCP server exists and is adequate | Reference it via its package manager distribution. We never fork-and-diverge. |
Contribute | A server exists but is partial | Contribute enhancements upstream, under the upstream project's license and governance. |
Build | No usable server exists | Build and maintain it here — vendor-neutral, permissively licensed, production-quality. |
Sunset policy: when a vendor ships an official server for a technology we Build, we converge — we contribute our learnings and tests upstream, put our server in maintenance mode with a migration note, and offer to donate the codebase to the vendor. The goal is ecosystem health, not ownership.
Related MCP server: Defender Hunt MCP
Coverage registry
Technology | Mode | Status | Server / upstream |
Microsoft security (Defender XDR, Defender for Endpoint, Entra ID, Purview Audit) | Build | Active | |
AbuseIPDB (IP reputation, blacklist) | Adopt | Active | Agents: AbuseIPDB's own |
Google SecOps (Chronicle) | Contribute | Planned | Official server exists but is partial; upstream contributions being scoped |
CrowdStrike Falcon | Adopt | Evaluating | Official/third-party options under evaluation |
Microsoft Sentinel | — | Candidate | Not yet scoped |
Want a technology covered? Open an issue with the vendor, the API surface, and links to any existing MCP servers.
One technology, two parts
Every integration built here ships as a tool client (clients/<technology>/, a typed Python library) and an MCP server (servers/<technology>/) that is a thin wrapper over it. Where the technology's own server is adopted, only the client lives here, for hosts that call the API from code (AbuseIPDB). Agents use the server from any MCP host; programs call the client directly where the path must be deterministic. Each API call is implemented once, so both see the same behaviour. Each server also ships a capabilities manifest and, where the technology allows it, a capability probe that writes a ZeroSOC capability binding file for what a given tenant exposes.
Servers run standalone. The server and its tool client are two packages in this repository, so both are named on the command line:
uvx --from "git+https://github.com/ZeroSOC/zerosoc-mcp@<tag>#subdirectory=servers/<technology>" \
--with "git+https://github.com/ZeroSOC/zerosoc-mcp@<tag>#subdirectory=clients/<technology>" \
<command>Quality bar
Every server built here meets the same bar:
Safe by default — read tools and response/write tools are separated; destructive response actions are disabled unless explicitly enabled per deployment via an environment flag.
Context-friendly — every list tool paginates with sane caps; no unbounded dumps into an agent's context window.
Actionable errors — API failures surface the status and a remediation hint (e.g. the exact missing permission).
No secrets — configuration via environment variables only, documented in each server's
.env.example; required API permissions are documented per tool group.
Servers
servers/microsoft-security — Microsoft security
Incidents with their full evidence inventory, alerts, cross-workload advanced hunting, Entra ID logs and the Microsoft 365 unified audit log (Microsoft Purview Audit) via Microsoft Graph, a capability probe, plus device response actions, threat indicators, and vulnerability management via the Microsoft Defender for Endpoint API. See its README for setup, and clients/microsoft-security for the tool client.
Contributing
See CONTRIBUTING.md. Contributions are accepted under Apache 2.0 with DCO sign-off (git commit -s); there is no CLA. Taking part means keeping to the Code of Conduct.
Please do not put real tenant data in an issue or a pull request — no device or user names, tenant identifiers, hashes or incident exports.
Security
Found a vulnerability? Do not open a public issue. See SECURITY.md for private reporting and what is in scope.
License
Apache 2.0. Portions derived from MIT-licensed prior work — see NOTICE.
Available Tools
66 toolsdefender_add_alert_commentA
Append a comment to an alert's comment thread (a triage write), visible to analysts in the Defender portal. Notes about the whole case belong on the incident instead.
Inputs: comment (required), alert_id (required).
defender_add_incident_commentA
Append a comment to an incident's comment thread (a triage write), visible to analysts in the Defender portal. If the incident was merged into another, the comment is written on the master incident, where analysts will read it, and redirectedFrom says so.
Inputs: comment (required), incident_id (required).
defender_add_remove_machine_tagA
Add a tag to a device or remove one (a metadata write).
Inputs: value (required), action (required), machine_id (required).
defender_decode_commandARead-only
Decode a PowerShell -EncodedCommand payload (base64 of UTF-16LE) exactly. Accepts the full
command line, any spelling of the flag (-e, -enc, -ec, ...), or the bare base64. Use it
instead of decoding by hand. A decoding that is itself an encoded command is decoded again,
up to five rounds: rounds is how many layers were removed and capped says the command was
still encoded when they ran out. The evidence inventory already carries decoded_command,
decode_rounds and decode_capped on process rows.
Inputs: command_line (required).
defender_export_antivirus_healthARead-only
Get the short-lived download links of the full antivirus health report of every device. For a bounded look at it use defender_get_device_health.
Inputs: no named parameters.
defender_export_assessmentARead-only
Read a per-device assessment: one row per device and software, vulnerability, configuration or browser extension. The full export is large; this returns one bounded page and the next link.
Inputs: assessment_type (required), top.
defender_find_machines_by_ipARead-only
Find the devices that had an IP address around a given time.
Inputs: ip (required), timestamp (required), top.
defender_find_machines_by_tagBRead-only
Find the devices that carry a tag.
Inputs: tag (required), top, starts_with.
defender_get_alertARead-only
Get one alert with its full evidence (files, processes, registry keys, IPs, URLs, users, mailboxes, devices), MITRE techniques, detection source, detector ID and comment thread.
Inputs: alert_id (required).
defender_get_capabilitiesARead-only
What this tenant exposes, as a ZeroSOC capability binding. Call it once before
investigating. It checks which hunting tables hold data, which are exposed but empty and
which the licence does not expose at all, which APIs answer, and which application
permissions are granted, so an empty query result is read correctly: a visibility gap, a
missing licence or a missing permission. Returns capabilities (capability class to tool),
data_sources (each playbook data source, available or not, with the reason in
data_source_notes) and probe (every check, the granted roles, the classes left unbound
and the checks that stay manual). The probe only reads; the result is kept until refresh.
Inputs: refresh.
defender_get_device_healthARead-only
List the antivirus health of devices: mode, engine, platform and signature versions and their freshness, last quick and full scan. A device whose antivirus is passive, disabled or out of date explains a missing detection.
Inputs: top, filter.
defender_get_domain_alertsBRead-only
List the Defender for Endpoint alerts that involve a domain.
Inputs: domain (required), top, filter.
defender_get_domain_machinesBRead-only
List the devices that communicated with a domain: the lateral scope of a domain.
Inputs: domain (required), top, filter.
defender_get_domain_statisticsARead-only
Get the prevalence of a domain in this organization, with first and last seen.
Inputs: domain (required).
defender_get_exposure_scoreARead-only
Get the organization's exposure score: lower is better.
Inputs: no named parameters.
defender_get_file_alertsARead-only
List the Defender for Endpoint alerts that involve a file.
Inputs: file_hash (required), top, filter.
defender_get_file_infoARead-only
Get what Defender knows about a file by hash: global prevalence and first seen, size, type, signer and issuer, determination. Only the hash leaves the environment.
Inputs: file_hash (required).
defender_get_file_machinesARead-only
List the devices a file was seen on: the lateral scope of a hash.
Inputs: file_hash (required), top, filter.
defender_get_file_statisticsARead-only
Get the prevalence of a file in this organization and worldwide, with first and last seen. A file seen on one device and nowhere else is worth a closer look.
Inputs: file_hash (required).
defender_get_incidentARead-only
Get one incident as it is stored: severity, status, classification, determination, owner, custom tags, comment thread, timestamps, and redirectIncidentId when it was merged into another. The starting point of incident analysis; it does not follow merges (use defender_resolve_incident for that).
Inputs: incident_id (required).
defender_get_incident_alertsARead-only
Get the alerts of an incident, paged (skip/top), with their details: title, severity, category, MITRE techniques, detection source, timestamps and nested evidence. For the entities themselves prefer defender_get_incident_evidence, which flattens and de-duplicates the evidence of every alert in one call.
Inputs: incident_id (required), top, skip, summary_only.
defender_get_incident_evidenceARead-only
The full evidence inventory of an incident: the portal's evidence view as one table. Pages
through every alert and flattens the evidence into one row per evidence item, each joined to
its device (by device ID, else through the alert) and listing the alerts that cite it. A row
is the source's own unit: entityCount equals the portal's count of evidence items and is the
number to reconcile; rowCount is the number of per-alert evidence rows. The source lists one
process more than once when alerts describe it differently (no image file, another path, a
different verdict): those rows share instance (device, PID, creation time), sameProcess
groups them with their verdicts, and processInstanceCount is the number of real processes.
Rows carry type, name, PID and parent PID, full command line and its decoded
-EncodedCommand payload, creation times in UTC, SHA1/SHA256, path, user SID and UPN, device
ID, IP, registry key and value, detailed roles, verdict and remediation status. Follows a
merged incident to its master. These are the entities the alerts cite, not raw telemetry:
processes the detection did not flag are not here.
Inputs: incident_id (required), top, skip, types.
defender_get_indicatorsARead-only
List the custom indicators (allow, audit, warn, block) of the tenant, with OData filtering, e.g. "indicatorValue eq '203.0.113.7'", "action eq 'Block'".
Inputs: top, skip, filter.
defender_get_investigation_by_idARead-only
Get one automated investigation: state, status details, device and triggering alert.
Inputs: investigation_id (required).
defender_get_investigationsARead-only
List automated investigations with their state and the alert that triggered each.
Inputs: top, skip, filter.
defender_get_ip_alertsARead-only
List the Defender for Endpoint alerts that involve an IP address.
Inputs: ip (required), top, filter.
defender_get_ip_statisticsARead-only
Get the prevalence of an IP address in this organization, with first and last seen.
Inputs: ip (required).
defender_get_live_response_resultARead-only
Get the short-lived download link of the output of one live response command.
Inputs: action_id (required), command_index (required).
defender_get_machine_action_by_idARead-only
Get one machine action: type, status (Pending, InProgress, Succeeded, Failed, TimeOut, Cancelled), requestor, comment and times. Poll it after a response action.
Inputs: action_id (required).
defender_get_machine_actionsARead-only
List response actions taken on devices (isolation, release, scans, quarantine, packages, live response) with their status, requestor, comment and times, e.g. filter "machineId eq ''" or "type eq 'Isolate'". It lists actions taken by anyone, people included, so it is how to check whether a containment still stands.
Inputs: top, skip, filter.
defender_get_machine_alertsARead-only
List the Defender for Endpoint alerts of one device.
Inputs: machine_id (required), top, filter.
defender_get_machine_by_idARead-only
Get one device: DNS name, OS, health, risk and exposure level, tags, device value, IPs, first and last seen. Accepts the machine ID (the mdeDeviceId of alert evidence).
Inputs: machine_id (required).
defender_get_machine_group_exposure_scoreBRead-only
Get the exposure score of each device group.
Inputs: no named parameters.
defender_get_machine_logon_usersARead-only
List the users that logged on to a device.
Inputs: machine_id (required), top.
defender_get_machinesARead-only
List devices onboarded to Defender for Endpoint, with OData filtering, e.g. "healthStatus eq 'Active'", "riskScore eq 'High'", "startswith(computerDnsName,'ws')".
Inputs: top, skip, filter.
defender_get_machines_by_softwareBRead-only
List the devices that have a software installed.
Inputs: software_id (required), top.
defender_get_machines_by_vulnerabilityBRead-only
List the devices exposed to a vulnerability.
Inputs: cve_id (required), top.
defender_get_machine_vulnerabilitiesARead-only
List the vulnerabilities that affect one device.
Inputs: machine_id (required), top.
defender_get_package_sas_uriARead-only
Get the short-lived download link of a collected investigation package. The link grants access to forensic data: hand it to a person, do not store it in a note.
Inputs: action_id (required).
defender_get_recommendation_by_idCRead-only
Get one security recommendation.
Inputs: recommendation_id (required).
defender_get_recommendation_machinesBRead-only
List the devices a security recommendation applies to.
Inputs: recommendation_id (required), top.
defender_get_recommendationsARead-only
List security recommendations, e.g. filter "remediationType eq 'ConfigurationChange'".
Inputs: top, skip, filter.
defender_get_recommendation_vulnerabilitiesARead-only
List the vulnerabilities a security recommendation addresses.
Inputs: recommendation_id (required), top.
defender_get_remediation_activitiesBRead-only
List remediation activities with their status and due date.
Inputs: top, filter.
defender_get_remediation_activity_by_idCRead-only
Get one remediation activity.
Inputs: activity_id (required).
defender_get_remediation_exposed_devicesARead-only
List the devices still exposed under a remediation activity.
Inputs: activity_id (required), top.
defender_get_secure_scoreARead-only
Get the Microsoft Secure Score for Devices: higher is better.
Inputs: no named parameters.
defender_get_softwareBRead-only
List the software inventory of the organization with weaknesses and exposed devices.
Inputs: top, skip, filter.
defender_get_software_by_idBRead-only
Get one software entry: vendor, weaknesses, exploit availability, exposed devices.
Inputs: software_id (required).
defender_get_software_version_distributionARead-only
Get the distribution of installed versions of a software.
Inputs: software_id (required).
defender_get_software_vulnerabilitiesBRead-only
List the vulnerabilities of one software entry.
Inputs: software_id (required), top.
defender_get_user_alertsBRead-only
List the Defender for Endpoint alerts that involve a user.
Inputs: user_id (required), top, filter.
defender_get_user_machinesARead-only
List the devices a user logged on to: logon relationships, not role or privilege.
Inputs: user_id (required), top, filter.
defender_get_vulnerabilitiesBRead-only
List the vulnerabilities that affect the organization, e.g. filter "severity eq 'Critical'" or "publicExploit eq true".
Inputs: top, skip, filter.
defender_get_vulnerability_by_idARead-only
Get one vulnerability by CVE ID: severity, CVSS, exploit availability, exposed devices.
Inputs: cve_id (required).
defender_list_alertsARead-only
List alerts from all Defender XDR workloads: Endpoint, Office 365 email, Identity, Cloud Apps and Entra ID Protection. Supports OData filtering with Graph camelCase values, e.g. "severity eq 'high'", "status eq 'new'", "serviceSource eq 'microsoftDefenderForEndpoint'", "createdDateTime gt 2026-01-01T00:00:00Z". Each alert carries its evidence inline. For the alerts of one incident use defender_get_incident_alerts.
Inputs: top, filter, orderby.
defender_list_incidentsARead-only
List incidents from Microsoft Defender XDR. Incidents group related alerts from all Defender workloads (endpoint, identity, email, cloud apps) into one attack story. Supports OData filtering with Graph camelCase values, e.g. "status eq 'active'", "severity eq 'high'", "lastUpdateDateTime gt 2026-01-01T00:00:00Z", "assignedTo eq 'analyst@contoso.com'". Use it to find incidents before drilling into one. A merged incident has status 'redirected' and names its master in redirectIncidentId.
Inputs: top, filter, orderby.
defender_list_library_filesARead-only
List the scripts in the live response library.
Inputs: top.
defender_resolve_incidentARead-only
Follow the merge history of an incident to its master. Returns the master incident with incidentId (the master's ID) and redirectedFrom (the merged IDs walked, in order; empty when the incident is live). Writes made by this server already do this; call it to learn where an incident went.
Inputs: incident_id (required).
defender_run_hunting_queryARead-only
Run an advanced hunting query (KQL) across Defender XDR data: endpoint, identity, email and cloud apps in one query. Tables include AlertInfo, AlertEvidence, DeviceProcessEvents, DeviceNetworkEvents, DeviceFileEvents, DeviceRegistryEvents, DeviceLogonEvents, DeviceEvents, IdentityLogonEvents, IdentityDirectoryEvents, EmailEvents, EmailUrlInfo, UrlClickEvents and CloudAppEvents. Which tables hold data depends on the tenant's licences: an empty result from a table that defender_get_capabilities reports as empty or not exposed is a visibility gap, not a finding and not a permissions problem. The sensor records no creation event for some processes (services started before it, some short-lived instances): a process absent from DeviceProcessEvents may still be there as the InitiatingProcessId and InitiatingProcessCreationTime of other events, so look there before concluding it did not run. Match a process by PID together with its creation time, never by PID alone. Timestamps are UTC. Data is limited to the last 30 days. Example: "DeviceProcessEvents | where Timestamp > ago(1d) | where FileName =~ 'powershell.exe' | project Timestamp, DeviceName, ProcessCommandLine | limit 50".
Inputs: query (required), timespan.
defender_to_utcARead-only
Convert a timestamp to UTC exactly. The APIs return UTC; the Defender portal shows the viewer's local time. Convert every time quoted from the portal or by a person before it goes into a timeline, so all entries share one reference.
Inputs: timestamp (required), assume_zone.
defender_update_alertA
Update an alert (a triage write): status, assignee, classification and determination.
Inputs: alert_id (required), status, assigned_to, determination, +1 more.
defender_update_incidentA
Update an incident (a triage write): status, assignee, classification, determination and custom tags. If the incident was merged into another, the update is applied to the master incident and redirectedFrom says so.
Inputs: incident_id (required), status, assigned_to, custom_tags, +2 more.
defender_update_machineB
Update a device's tags or device value (a metadata write).
Inputs: machine_id (required), device_value, machine_tags.
entra_list_directory_auditsARead-only
List Entra ID directory audit events: changes to users, groups, roles, applications, service principals, credentials and policies, with who made them. Always filter, e.g. "activityDisplayName eq 'Add member to role'", "initiatedBy/user/userPrincipalName eq 'alice@contoso.com'", "activityDateTime ge 2026-01-01T00:00:00Z", "targetResources/any(t: t/id eq '')". Timestamps are UTC.
Inputs: top, filter, orderby.
entra_list_sign_insARead-only
List interactive sign-ins from the Entra ID sign-in log: user, application, client and IP address, location, device, conditional access result, MFA detail, risk and error code. Always filter, e.g. "userPrincipalName eq 'alice@contoso.com' and createdDateTime ge 2026-01-01T00:00:00Z", "ipAddress eq '203.0.113.7'", "status/errorCode ne 0". Timestamps are UTC. Requires an Entra ID P1 or P2 licence in the tenant; retention is 30 days.
Inputs: top, filter, orderby.
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 66 tools
Most tools use a clear resource-plus-attribute pattern (get_machine_*, get_file_*, get_recommendation_*), so their purposes are largely distinct. A few inverse/overlapping pairs—machine_vulnerabilities vs machines_by_vulnerability, file_info vs file_statistics, incident_alerts vs incident_evidence—require careful reading and could cause misselection.
The dominant defender_get_<resource> and defender_update_<resource> conventions are consistent and readable. Deviations are minor but visible: mixed get_/list_ for collections, find_machines_by_* instead of get_*, the combined add_remove verb, and an entra_ prefix inside a defender_ server.
66 tools is far beyond the typical well-scoped MCP surface and exceeds the 50+ extreme-mismatch boundary. Even though the tools are grouped by domain, this scale creates significant selection and maintenance overhead and would be better split into focused servers.
Investigation, hunting, alert/incident triage, and vulnerability views are well covered. However, the response half of XDR is missing: there are no tools to isolate/release machines, run live response, create/delete indicators, or mutate remediation activities, leaving obvious dead ends after detection.
Maintenance
Related MCP Connectors
Deltabridge brings an AI-native cyber threat intelligence platform into your AI agent. Search a knowledge graph of 400,000+ threat actors, malware families, vulnerabilities, indicators and campaigns, mapped to MITRE ATT&CK. Look up and enrich IOCs, trace links between entities and read living briefs built from 300+ open sources. Read-only, scoped to your organisation, with OAuth sign-in or an API key. Included on Researcher, Business and Enterprise plans.
Lets AI agents analyze API metrics, request logs, and traces, and manage apps and alerts.
Governed app access for AI agents: 1,000+ apps & 12,000+ tools via Code Mode MCP.
Governed access to sensitive data, secrets and APIs for AI agents. A signed proof for each decision.
Related MCP Servers
- FlicenseNot gradedqualityDmaintenanceEnables agentic SecOps by executing Microsoft Defender XDR response actions (device isolation, antivirus scans, forensic collection, incident management) through natural language with orchestrators like GitHub Copilot.22-
- AlicenseNot gradedqualityDmaintenanceEnables security investigation and threat hunting through Microsoft Defender and Entra ID, with 31 tools for KQL queries, alerts, threat intelligence, identity investigation, and advanced threat hunting.MIT
- AlicenseAqualityBmaintenanceEnables security analysts to investigate Microsoft Entra ID security logs through natural language, exposing read-only tools for user context, sign-ins, risky users, risk detections, directory audits, and conditional access policies.6MIT
- AlicenseNot gradedqualityBmaintenanceEnables AI assistants to query Microsoft Defender XDR telemetry for read-only threat hunting, incident triage, and vulnerability discovery using delegated per-user authentication and production guardrails.2MIT