Azure NeuroMap MCP
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., "@Azure NeuroMap MCP map my Azure estate and show what's publicly exposed"
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.
Azure NeuroMap MCP
A read-only MCP server that maps your Azure estate as a graph and tells you, precisely and with evidence, what is exposed, how it is connected, and who can access it.
Every answer cites the exact Azure fields it came from. If something cannot be read, it says unknown and why. It never guesses. See the demo map (synthetic data).
A read-only MCP server that scans Azure and turns every resource into a connected graph: networking (VNets, subnets, NSGs and every rule, NICs, IPs, peerings, routes, gateways, firewalls), data and PaaS (SQL, Managed Instance, PostgreSQL, MySQL, Cosmos DB, Storage, Key Vault, App Service and Functions, AKS, ACR, Redis, Service Bus, Event Hubs, AI services, Search), edge devices (Azure Firewall and policies, Load Balancer, Application Gateway and WAF, VPN / ExpressRoute), managed identities and Azure RBAC. For each resource it states its network access and auth settings precisely, with the source of every fact.
Precision contract
Every fact records its source:
arg(Resource Graph) orarm <api-version>(direct GET).Every verdict lists the exact fields and values it was derived from (
because).Missing or unreadable data gives
unknownwith the reason (for exampleARM read failed (HTTP 403)).Where Azure documents what an unset field means (Key Vault without
networkAclsallows all networks, SQL with no firewall rules denies all), that rule is applied and written intobecause.Nothing is inferred from names, tags, naming conventions or connection strings.
No secrets are read: no
listKeys, no app settings, no connection strings. Only GET and Resource Graph.Rules are shown even when they do not apply. A resource with public access disabled still lists its firewall, IP and VNet rules, marked
rules_in_effect: false, because they come back the moment access is re-enabled. No allow edges are drawn for them.A "disable" flag that is not set is reported as not disabled (for example
local_auth_disabled), with the reason innotes.
Related MCP server: AWS Cost Saver
How it works
Azure Resource Graph (1 KQL query, paged) -> raw snapshot -> graph builder -> MCP tools
| |
~/.neuromap/snapshots HTML "neurosystem" mapOnly two Azure calls exist, both read-only: list subscriptions, and query Resource Graph. The raw snapshot is kept so later phases can re-derive new views without re-scanning.
Install
Windows (use one fixed path, so the AI client config never changes on upgrade):
python -m venv C:\Tools\neuromap\.venv
C:\Tools\neuromap\.venv\Scripts\python.exe -m pip install git+https://github.com/KanenasCS/azure-neuromap-mcp
C:\Tools\neuromap\.venv\Scripts\neuromap.exe selftest # offline, no Azure neededLinux / macOS:
python3 -m venv ~/.neuromap-venv
~/.neuromap-venv/bin/pip install git+https://github.com/KanenasCS/azure-neuromap-mcp
~/.neuromap-venv/bin/neuromap selftestFirst scan
az login --tenant <tenant-id> # or managed identity / AZURE_* env vars
neuromap scan -s <subscription-id> # read-only; prints a summary
neuromap map # offline HTML map, prints the path
neuromap rebuild # re-derive the graph from the saved scan (no Azure calls)Permission needed: Reader on the subscriptions you scan. In the summary check
enrichment_error_count: 0 and collection_errors: [].
Connect an AI client
Claude Desktop. Add to claude_desktop_config.json:
{ "mcpServers": { "azure-neuromap": {
"command": "C:\\Tools\\neuromap\\.venv\\Scripts\\neuromap-mcp.exe" } } }On Windows the Claude Desktop installer is an MSIX package, and the file it reads is
%LOCALAPPDATA%\Packages\Claude_<id>\LocalCache\Roaming\Claude\claude_desktop_config.json
(not only %APPDATA%\Claude\...). Quit Claude Desktop from the tray before editing, and
check Settings > Developer shows the server as running.
VS Code (.vscode/mcp.json):
{ "servers": { "azure-neuromap": { "type": "stdio",
"command": "C:\\Tools\\neuromap\\.venv\\Scripts\\neuromap-mcp.exe" } } }Upgrade
Quit the AI client first (Windows cannot replace a running .exe), then:
C:\Tools\neuromap\.venv\Scripts\python.exe -m pip install --upgrade --no-deps --force-reinstall git+https://github.com/KanenasCS/azure-neuromap-mcpSaved graphs from an older version are rebuilt from the raw scan automatically.
Tools
Tool | What it answers |
| Scan now: all resources, RGs, RBAC, access settings ( |
| Counts by kind, public-endpoint states, RBAC status, enrichment errors |
| Name, partial name, resource ID or IP to node, plus the subnet that contains an IP |
| One resource plus everything linked to it (depth 1 to 3) |
| Public-endpoint verdict, |
| Resources by verdict (default: open to all networks) |
| RBAC role assignments applying to a resource, direct and inherited (resource, RG, subscription, management groups, root), per-scope coverage (uncollected = unknown, never zero), capabilities from each role's actions/notActions (assign roles, write, delete), plus data-plane auth |
| Roles held by a resource's system and user-assigned identities |
| Azure Firewall or policy rules in processing order, parent policy inheritance, IP groups expanded, threat intel / IDPS / DNS proxy |
| Search all firewalls by port, source, destination, FQDN (wildcards), action, rule type |
| Every declared way in: firewall DNAT, public LB rules and NAT rules, App Gateway routes with effective WAF, public IPs on NICs |
| UDRs for a subnet/NIC/VM/VMSS with each next hop resolved to the resource that owns the IP |
| VPN/ER gateways, connections and status, on-prem ranges, ER peerings, P2S pools, spokes using the gateway |
| NSG layers on a VM/NIC/PE/public IP/subnet (or the subnet of AKS, MI, ...) in evaluation order |
| Estate-wide NSG rule search, e.g. port 3389 + source |
| Writes the offline interactive map and returns the path |
Prompts: map_my_infra, explain_resource, access_review, internet_ingress_review. Resource: neuromap://summary.
Access verdicts
public_endpoint is one of disabled, vnet_injected, restricted, all_networks,
all_networks_with_denies, enabled_no_allow_rules, gated_by_nsg, perimeter_controlled,
no_inbound_endpoint, decided_per_app, unknown, not_evaluated.
Type | Fields evaluated | Extra ARM reads |
Storage | publicNetworkAccess, networkAcls (default action, IP, VNet, resource rules, bypass), shared key, blob public access, TLS | none |
Key Vault | publicNetworkAccess, networkAcls, RBAC mode, access policies | none |
SQL server | publicNetworkAccess, firewall rules (union checked for full IPv4), 0.0.0.0 Azure rule, VNet rules (shown but marked not in effect when public access is disabled), Entra admin and Entra-only from their own child resources, TLS | server, firewallRules, virtualNetworkRules, administrators, azureADOnlyAuthentications (2021-11-01) |
SQL Managed Instance | publicDataEndpointEnabled, subnet, proxy, Entra-only, TLS | none |
PostgreSQL / MySQL flexible | delegated subnet, publicNetworkAccess, firewall rules, Entra/password auth | firewallRules (2022-12-01 / 2021-05-01) |
Cosmos DB | publicNetworkAccess, IP rules, VNet filter and rules, bypass, local auth, TLS | none |
App Service / Functions | publicNetworkAccess, access restrictions (first-match, default action), SCM site, VNet integration, FTP/SCM basic auth, TLS, FTPS | config/web, basicPublishingCredentialsPolicies (2023-01-01) |
AKS | private cluster, authorized IP ranges, node subnets, Entra/Azure RBAC, local accounts, network plugin/policy | none |
ACR | publicNetworkAccess, network rule set, SKU rule, admin user, anonymous pull | none |
Redis | publicNetworkAccess, VNet injection, firewall rules, access-key auth, non-SSL port, TLS | firewallRules (2022-06-01) |
Service Bus / Event Hubs | publicNetworkAccess, network rule set, trusted services, local auth, TLS | networkRuleSets/default (2021-11-01) |
AI services / OpenAI | publicNetworkAccess, networkAcls, local auth, outbound restriction | none |
AI Search | publicNetworkAccess, IP rules, local auth | none |
Managed disks | publicNetworkAccess, networkAccessPolicy (AllowAll / AllowPrivate / DenyAll), disk access, live export SAS ( | none |
Container Apps | Environment internal/external (internal overrides any app's external ingress), infrastructure subnet, app ingress external, IP restrictions (all-Allow or all-Deny semantics), insecure HTTP, client certs | none |
Logic Apps (Consumption) | Inbound triggers in the definition (Request, ApiConnectionWebhook, HttpWebhook), with trigger types cited, accessControl.triggers.allowedCallerIpAddresses ( | none |
Log Analytics / App Insights | publicNetworkAccessForIngestion and ForQuery (verdict = most open endpoint, both cited), local auth | none |
Data Collection Endpoints | networkAcls.publicNetworkAccess (Enabled / Disabled / SecuredByPerimeter) | none |
Automation | publicNetworkAccess (boolean) for webhooks and Hybrid Worker endpoints, local auth | none |
Purview | publicNetworkAccess, managed resources public access | none |
Any other type | raw publicNetworkAccess and private endpoint connections, marked | none |
Edge devices
Device | What is turned into facts and edges | Extra ARM reads |
Azure Firewall | Policy chain (base policies), rule collection groups, DNAT/network/application rules in processing order, IP groups expanded, classic rule collections, threat intel, IDPS, DNS proxy, vWAN hub IPs. DNAT rules become | ruleCollectionGroups (2023-09-01) |
Load Balancer | Frontends, backend pools (NIC, VMSS and IP-based members, deduplicated when Azure lists a node twice), owning AKS cluster and its power state (from the cluster's nodeResourceGroup), LB rules (incl. HA ports, floating IP), inbound NAT rules (single port and port-range v2), outbound rules, probes. Each rule becomes | none |
Application Gateway | Listeners (port, protocol, host names), basic and path-based routing, redirects, backend pools (NIC, IP, FQDN matched to web app and Container App host names), HTTP settings, SSL policy. WAF per route, resolved in order: path rule policy, listener policy, gateway policy, legacy WAF config, and "not available" when the SKU tier is not WAF | none |
Route tables | Each | none |
VPN / ExpressRoute | Gateway type, SKU, active-active, BGP, point-to-site pool and auth, S2S / VNet-to-VNet / ER connections with status and BGP, local network gateway on-prem ranges, ER circuit provider and peerings, spokes that use the gateway through peering | none |
VM Scale Sets | Subnet, NIC-configuration NSG, LB and App Gateway pool membership, instance public IPs flag | none |
Anything referenced but not owned by a scanned resource becomes an explicit node such as
ip:10.9.9.9 or fqdn:api.example.com, marked outside scan, never matched by guesswork.
Connection shared keys and certificates are never read.
A verdict describes the resource's own network layer. It does not claim a given client can connect end to end, because client-side DNS, routes, NSGs and firewalls are separate layers.
Relationship types
Network: CONTAINS IN_SUBNET PROTECTED_BY HAS_NIC HAS_PUBLIC_IP PEERED_WITH
ROUTES_VIA EGRESS_VIA CONNECTS_TO MEMBER_OF BALANCES_TO USES_POLICY
Access: OPEN_TO_ALL_NETWORKS (Internet node to resource), IP_RULE_ALLOWS (a scanned public IP
is inside an allow rule), VNET_RULE_ALLOWS (subnet allowed by a VNet rule). Allow edges are
drawn only for restricted resources, where they actually change who can connect.
Platform and identity: VNET_INTEGRATION HOSTED_ON USES_IDENTITY MANAGED_BY HAS_ROLE
(one edge per role assignment).
Edge: DNAT_FORWARDS LB_FORWARDS APPGW_ROUTES NEXT_HOP HYBRID_CONNECTION
INGRESS_ANY_SOURCE INHERITS_FROM USES_IP_GROUP ATTACHED_TO IN_HUB
Generic: REFERENCES for any other resource ID written in a resource's properties, with the
exact property path (for example properties.parameters.$connections.value.sql.connectionId).
Data handling
Snapshots and maps describe your network in detail. They live in ~/.neuromap
(change with NEUROMAP_HOME). Treat that folder as sensitive and do not commit it.
The map HTML embeds Cytoscape.js (MIT, see static/cytoscape.LICENSE) and makes no
network calls.
Known limits
RBAC: assignments above the subscription are read with
roleAssignments?$filter=atScope()(Reader is enough); if that call fails, those scopes are reported as not collected. Active assignments only. Deny assignments, PIM eligible roles and group membership are not expanded, and users/groups/service principals are shown by object ID (no Microsoft Graph calls).Networking is declared configuration. Service tags and ASG membership are not expanded, and routes and firewalls are not evaluated for end-to-end reachability.
Not evaluated yet: Azure-computed effective routes and system/BGP routes (needs the effectiveRouteTable action, which Reader cannot call), NSG flow logs, Front Door, Traffic Manager, Virtual WAN routing intent, App Gateway rewrite rules and per-URI WAF policies on path maps beyond the path rule level.
ARM enrichment makes a few GETs per SQL server, web app, flexible server, Redis cache and Service Bus/Event Hubs namespace. Failures become
unknown, never guesses.Map rendering takes around 15 seconds at 8,000+ nodes. Use the legend to hide noisy kinds.
Validation. 151 offline checks (synthetic fixtures shaped like Resource Graph and ARM responses, a mocked Azure API for the collector, and order-independence tests) run on Linux and Windows, Python 3.10 to 3.14. Every evaluator was also compared field by field with the Azure CLI on a live estate of 100+ resources; the fixes that produced are in CHANGELOG.md.
Roadmap
Phase 2, security overlay: Defender for Cloud plan coverage and recommendations, DDoS, WAF mode, NSG flow logs, diagnostic settings, per node.
Phase 3, reasoning: chain the declared layers (DNAT or LB, then UDR, then firewall rule, then NSG) into end-to-end path verdicts, blast radius, and snapshot diffs ("what changed since last week").
Security and contributing
Read SECURITY.md before your first scan. Scans describe your environment in detail and must never be shared. Contributions follow the precision contract in CONTRIBUTING.md.
License
MIT. Cytoscape.js is vendored under its own MIT license (src/neuromap/static/cytoscape.LICENSE).
Available Tools
16 toolsexport_mapARead-only
Write the interactive offline 'neurosystem' HTML map of the current snapshot and return its path.
Inputs: no named parameters.
find_resourceARead-onlyIdempotent
Find nodes by full resource id, IP address (private or public), exact or partial name.
kind filters results: network kinds (vnet, subnet, nsg, nic, vm, vmss, pip, pe, lb, appgw, firewall, fwpolicy, wafpolicy, ipgroup, vnetgw, lng, ercircuit, gwconnection, ...), data/PaaS kinds (storage, keyvault, sql-server, sql-db, sql-mi, postgres, mysql, cosmos, webapp, aks, acr, redis, servicebus, eventhub, cognitive, search, uami), or the short type of any other resource (e.g. 'logic/workflows'). For an IP, declared ranges containing it are also returned (subnets, on-premises ranges from local network gateways, point-to-site pools).
Inputs: query (required), kind, limit.
firewall_rulesBRead-onlyIdempotent
Azure Firewall (or Firewall Policy) rules in processing order: DNAT, network, application. Parent-policy rule collection groups come first. IP groups are expanded. Includes threat intel, IDPS and DNS proxy settings, and classic (non-policy) rules. Missing policy data is listed as unknown.
Inputs: target (required).
get_accessBRead-onlyIdempotent
Network access verdict for one resource (SQL, Storage, Key Vault, Cosmos, App Service, AKS, ACR, Redis, PostgreSQL, MySQL, Service Bus, Event Hubs, AI services, Search, ...).
Returns public_endpoint (disabled | vnet_injected | restricted | all_networks | ...),
because (exact fields used), network rules, allowed IPs/subnets, private endpoints,
auth settings (Entra-only, local auth, shared key), TLS, per-fact sources, and unknowns.
Inputs: target (required).
get_nodeBRead-onlyIdempotent
Full detail of one node plus everything linked to it, up to depth 3.
ref: resource id, IP or unique name. Relationships include CONTAINS, IN_SUBNET, PROTECTED_BY, HAS_NIC, HAS_PUBLIC_IP, PEERED_WITH, ROUTES_VIA, EGRESS_VIA, CONNECTS_TO, MEMBER_OF, BALANCES_TO, USES_POLICY, VNET_INTEGRATION, HOSTED_ON, USES_IDENTITY, MANAGED_BY, HAS_ROLE, OPEN_TO_ALL_NETWORKS, VNET_RULE_ALLOWS, IP_RULE_ALLOWS, and REFERENCES (any other ARM id found in properties, with its exact property path).
Inputs: ref (required), depth.
hybrid_connectivityBRead-onlyIdempotent
VPN and ExpressRoute: each gateway with its VNet, SKU, BGP and point-to-site settings, its connections (type, status, BGP) and remote side (on-prem ranges, gateway IP, ER peerings), and the spoke VNets that use it through peering.
Inputs: no named parameters.
identity_permissionsCRead-onlyIdempotent
What a resource's system-assigned and user-assigned managed identities hold: roles and scopes.
Inputs: target (required).
infra_summaryBRead-onlyIdempotent
Counts by resource kind and relationship, scan time, subscriptions, unattached NSGs and public IPs.
Inputs: no named parameters.
ingress_pathsARead-onlyIdempotent
Every declared ingress path: firewall DNAT (with allowed sources), public load balancer rules and inbound NAT rules, App Gateway listener -> backend routes (with effective WAF mode), and public IPs attached directly to NICs.
target: limit to paths into or through one resource (VM, NIC, firewall, LB, App Gateway, IP). port: limit to a frontend port. include_private: also show internal LB / private App Gateway listeners.
Inputs: port, target, include_private.
list_public_exposureARead-onlyIdempotent
List resources by public-endpoint verdict. Default: all_networks and all_networks_with_denies.
states: any of disabled, vnet_injected, restricted, all_networks, all_networks_with_denies, enabled_no_allow_rules, gated_by_nsg, perimeter_controlled, no_inbound_endpoint, decided_per_app, unknown, not_evaluated. kind: e.g. storage, keyvault, sql-server, webapp, cosmos, aks, acr, redis, disk, containerapp, aca-env, logicapp, loganalytics, appinsights, dce, automation, purview.
Inputs: kind, limit, states.
nsg_rules_forARead-onlyIdempotent
Every NSG layer on a VM, NIC, private endpoint, public IP or subnet, in Azure evaluation order.
Inbound shows subnet NSG then NIC NSG. Outbound shows NIC NSG then subnet NSG. Rules are sorted by priority, default rules included.
Inputs: target (required).
route_forBRead-onlyIdempotent
User-defined routes that apply to a subnet, NIC, VM, VMSS or subnet-injected service, with each VirtualAppliance next hop resolved to the resource that owns that IP (firewall, NVA, internal LB), or to an explicit 'ip:x' node when no scanned resource owns it.
Inputs: target (required).
scan_infrastructureARead-onlyIdempotent
Scan Azure (read-only) and rebuild the graph: every resource, resource groups, RBAC role assignments, and access settings.
Uses DefaultAzureCredential (az login, managed identity, env vars). Needs Reader. subscription_ids: limit the scan; omit to scan every enabled subscription visible. enrich: also read child settings Resource Graph lacks (SQL/PostgreSQL/MySQL/Redis firewall rules, App Service access restrictions, Service Bus/Event Hubs network rules). Without it those verdicts are 'unknown'.
Inputs: enrich, subscription_ids.
search_firewall_rulesBRead-onlyIdempotent
Search every firewall's effective rules.
port: destination port (ranges and '*' match). source/destination: IP, CIDR, service tag or 'internet' (matches *, 0.0.0.0/0). fqdn: matched against target FQDN patterns (wildcards honored). action: Allow, Deny or DNAT. rule_type: dnat, network or application.
Inputs: fqdn, port, limit, action, +3 more.
search_nsg_rulesARead-onlyIdempotent
Search NSG rules across the whole estate, with what each NSG is attached to.
port: destination port, matched against single ports, ranges and '*'. source/destination: a CIDR or IP (overlap match), a service tag like 'VirtualNetwork', or 'internet' to match Internet, *, 0.0.0.0/0 and ::/0. access: Allow or Deny. direction: Inbound or Outbound. protocol: Tcp, Udp, Icmp.
Inputs: port, limit, access, source, +4 more.
who_has_accessBRead-onlyIdempotent
Active Azure RBAC role assignments that apply to a resource: on the resource, its parent resources, its resource group, subscription, management groups and root, marked inherited or direct. scope_coverage says per scope whether assignments were collected (a scope that was not collected is unknown, never zero). Each assignment has capabilities computed from the role definition's actions/notActions: assign_roles, write_resource, delete_resource. Also returns data-plane auth facts (Key Vault access policies, SQL Entra admin, ...).
Inputs: target (required).
Tool Schema Changelog
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
TDQS
Scored across 16 tools
Most tools have distinct scopes (find/get, scan/summary/export, per-resource vs estate-wide searches). A few pairs overlap conceptually—nsg_rules_for vs search_nsg_rules, firewall_rules vs search_firewall_rules, and who_has_access vs identity_permissions—but descriptions clarify per-resource vs global/bulk and RBAC vs managed-identity scopes.
All names use lowercase snake_case consistently, which is the dominant pattern. Minor deviations exist: some are verb-first (scan_infrastructure, find_resource, get_node), others noun-first (infra_summary, firewall_rules, ingress_paths), and who_has_access is a question phrase, but the set remains readable and predictable.
16 tools is slightly above the ideal 3–15 range but still well-scoped for a comprehensive Azure network graph and security analysis server. Each tool appears to cover a distinct analysis need, with no obvious filler.
The surface covers scanning, resource lookup, NSG/firewall rule analysis, routes, hybrid connectivity, RBAC/identity permissions, access verdicts, public exposure, and map export—strong coverage for read-only infrastructure analysis. Minor gaps like snapshot comparison or private DNS zone specifics are not fatal but leave some adjacent lifecycle concerns uncovered.
Related MCP Connectors
Cloud cost, inventory and governance on AWS/Azure/GCP. Read-only by default, optional scoped writes
CTEM for your Trusteed tenant: security summary, findings, compliance gaps, and scans via OAuth.
Read-only CVE intelligence, remediation playbooks, and agent setup guides. Not a scanner.
IaC attack-path auditor: finds internet-to-crown-jewel chains in Terraform/CFN/K8s.
Related MCP Servers
- FlicenseDqualityDmaintenanceEnables read-only assessment of AWS environments by inventorying resources, running security and operational checks, and generating actionable reports with cost analysis. Designed for contractors with support for assume-role authentication using external IDs.10-
- AlicenseNot gradedqualityDmaintenanceEnables scanning AWS accounts for cost optimization opportunities by reading resource configurations and pricing data, without making any modifications.24MIT
- AlicenseAqualityDmaintenanceEnterprise-grade Azure security assessment toolkit with multi-location scanning, IMDS exploitation, attack path analysis, and compliance reporting. Enables authorized penetration testing and compliance audits across all Azure regions.4311 npmMIT
- FlicenseNot gradedqualityBmaintenanceEnables AI to scan AWS accounts, analyze attack paths, and verify security fixes on a read-only graph of cloud resources.5-