One Config Line, One New Capability
Adding an MCP server to Claude or Cursor takes a minute. From then on the agent can read Slack, push code, or run shell commands, and no change ticket records it.
Secure MCP & AI Supply Chain
Control What Your AI Agents Can Do
Netra inventories the MCP servers, Agent Skills, local models, and AI components on every endpoint, scores the risky capability chains they create, and lets you audit or block the tools and skills you don’t trust.
Get a DemoASSETS › MCP
MCP
Create control policy
ToolAI ClientUsersDevicesRiskControl
exec_commandshell-runner-mcpCodex33HighBlocked
push_branchrepo-sync-mcpCursor1416HighAudit only
export_contactscrm-export-mcpChatGPT55HighBlocked
read_channelslack-readerClaude2225MediumAudit only
search_docsdocs-search-mcpGemini3133LowNot controlled
Product views on this page use illustrative data.
The Challenge
Agents Gain New Powers Without a Ticket
MCP servers, skills, and models install in minutes, on the endpoint.
Skills Ship Their Own Scripts
An Agent Skill is a SKILL.md plus files that can run commands, read the filesystem, and call the network. A same-name copy or an edited script looks identical from outside.
The Risk Is in the Combination
Reading labelled files is routine. Network access is routine. An agent that can do both on the same device is an exfiltration path, and a tool list won’t show it.
Inventory & Control
Four Layers of the Agentic Supply Chain
Found from what is configured, installed, and loaded on each endpoint.
01MCP Servers & Tools
Audit or Block Any MCP Tool
Assets › MCP shows every MCP tool configured across Claude, Cursor, ChatGPT, Codex, and Gemini, with users, devices, risk, and control state. Select tools and create a control policy: Audit only or Block, for all users or specific departments and people. MCP Activity records each call under Observe and Block.
ASSETS › MCP › CREATE CONTROL POLICY
1 · Tools2 · Action3 · Audience
Toolsshell-runner-mcp · exec_commandcrm-export-mcp · export_contacts
ActionAudit onlyBlock
Audience
All users Specific departments or users SalesFinanceM. Chen
02Agent Skills
Know What Each Skill Can Execute
Assets › Skills finds Agent Skills (SKILL.md) in Claude Code, Codex, and Claude Desktop, tags their capabilities (command execution, code execution, filesystem, network, data processing), and rates their risk. In whitelist mode you Allow or Revoke skills one by one, or you blacklist a skill by name or by exact build. Skill Activity logs each call, including the ones policy refused.
ASSETS › SKILLS
InstalledReview
Policy modeBlacklistWhitelist
SkillCapabilitiesRiskAction
deploy-helperClaude CodeCommand ExecutionNetworkHighAllow
csv-cleanerCodexCode ExecutionData ProcessingMediumRevoke
meeting-notesClaude DesktopFilesystemLowRevoke
ANALYTICS › SKILLS Shadow Skill Signals
Same-name impersonation1
Scripts changed2
03Local Models
Local Models, Proven by What Was Loaded
Assets › Models builds the local model inventory from real weight-load events in runtimes like Ollama and LM Studio. Each model shows its source registry, weights read, runtime, the endpoints and operators that loaded it, and evidence IDs. Model Loads adds the process and path behind each load, with flags such as user-scope installs.
ASSETS › MODELS
ModelSource registryWeights readRuntimeEndpointsEvidence
coder-7b-q4hub.models.example4.1 GBOllama6EV-20417
summarizer-3bregistry.llm.example2.0 GBLM Studio2EV-20388
vision-11bhub.models.example7.8 GBOllama1EV-20311
ACTIVITY › MODEL LOADS ModelRuntimeLoaded byFlag
vision-11bOllamapython3 · ~/tools/run.pyUser-scope install
04AI Components
The AI BoM Under Your Agents
AI BoM finds AI components that never show up as an app: npm and pip packages, VS Code and JetBrains extensions, Windows services, scheduled tasks, autorun registry keys, Ollama and LM Studio models, and WSL packages. Persistence points sit in the same list as packages and extensions. For AI apps and accounts, see Discover Shadow AI.
ASSETS › AI BOM
Source
All31
Packages12
IDE / Editor5
App / Browser2
System / Autorun7
AI runtime3
Linux / Unix2
NameSourceGroup
mcp-bridge-clinpm packagePackages
agent-toolkitPython pip packagePackages
ai-review-helperVS Code extensionIDE / Editor
agent-launcherScheduled taskSystem / Autorun
summarizer-3bLM Studio modelAI runtime
llm-toolsWSL packageLinux / Unix
Investigation
See Capabilities as Chains, Not Lists
The Netra AI Data Graph scores what each user’s agents can do together.
INCIDENTS › RISKY CHAINS
Agent Graph
Who runs what, with which model, through which tools, and what data it touched
Risky ChainsEntity 360
Agent ChainsSaaS ChainsShadow AI Domains
J. Lin → Claude CodeEngineering · PC-WIN-07 execnetworkfilesystemCritical capability14 MIP-labelled files on device 9.2 Critical
UserAgentCapabilitiesRisk score
J. LinEngineeringClaude Codeexecnetworkfilesystem9.2 Critical
S. PatelEngineeringCursornetworkfilesystem6.8 High
M. ChenFinanceCodexfilesystem4.4 Medium
Risk score 0–10CriticalHighMedium
Capability Composition
Agent Chains add up what each user’s agents can do across their devices: exec + network + filesystem, critical capabilities, and MIP-labelled files on the same device.
Entity 360
One graph of users, devices, agent channels, models, MCP servers, skills, data resources, and domains, linked by edges such as configures_mcp, has_skill, and accessed.
MCP Flow
Trace department → MCP server → tool → outcome. The Installed vs used view shows servers nobody calls, which is attack surface you can remove.
Next Steps
Secure the Rest of Your AI Use
The same endpoint agent and console cover data, accounts, and spend.
Secure AI Coding
Keep source code and secrets out of coding-agent prompts and responses.
Learn moreDiscover Shadow AI
See which AI apps, agents, and accounts your workforce actually uses.
Learn moreStop AI Data Leakage
Detect and block sensitive data in prompts, uploads, and responses.
Learn moreGovern AI Spend
Attribute AI token spend to users and departments, then cap it.
Learn moreFAQ
Common Questions
What does the AI supply chain on an endpoint include?
Everything an AI agent can load or call on an employee’s device: MCP servers and their tools, Agent Skills, local models, and AI components installed through package managers, IDEs, and system services. Netra inventories each one and ties it to users and devices.
Which AI clients does Netra find MCP servers in?
Assets › MCP lists every MCP tool configured across Claude, Cursor, ChatGPT, Codex, and Gemini, with the users and devices that have it, a risk level, and a control state of Audit only, Blocked, or Not controlled.
Can Netra block a specific MCP tool or Agent Skill?
Yes. Create a control policy for selected MCP tools with the action Audit only or Block, for all users or specific departments or users. For Agent Skills, use whitelist mode to Allow or Revoke each skill, or blacklist a skill by name or by exact build. Endpoints quarantine a blacklisted skill on their next policy sync, usually within a minute.
How does Netra score risky agent chains?
Incidents › Risky Chains builds an Agent Graph of who runs what, with which model, through which tools, and what data it touched. Each chain gets a risk score from 0 to 10 and a critical, high, or medium level, so the riskiest combinations of capability and data rise to the top.
Map Your Agentic Supply Chain
Run a proof of value on your own endpoints and see which MCP servers, skills, and models your agents can reach, and which chains need attention.
Get a Demo