MCP server
Give your AI assistant real dependency context
Connect an MCP-compatible assistant to the inventory and findings in your ShieldedStack workspace. Ask about affected projects, vulnerabilities, and package risk using read-only tools hosted by your on-premises Portal.
Ask questions against the workspace you actually run
An assistant can use ShieldedStack’s Model Context Protocol server to investigate the package versions and findings recorded in your workspace. That gives it project context for dependency questions instead of relying only on general model knowledge.
- “Which projects use lodash in this workspace?”
- “Show high and critical vulnerabilities affecting our npm packages.”
- “Which packages in the Backend project have the highest risk scores?”
- “Which NuGet packages are far behind their latest release?”
These are example questions for a connected assistant. Answers depend on the inventory and intelligence available in the selected workspace.
Four tools for dependency investigation
- find_vulnerabilities
- Search advisories by ecosystem, project, severity, or text.
- find_risky_packages
- Find packages ranked by risk score, advisory count, and CVSS.
- find_outdated_packages
- Identify deprecated packages or versions behind newer releases.
- get_project_inventory
- Find package usage across projects in the workspace.
Connect to your on-premises Portal
The MCP server is included in the Portal and uses Streamable HTTP at /mcp. Configure your MCP-compatible client with your Portal URL and bearer API-key authentication.
Enable Allow read-only workspace data access on a dedicated key. A normal proxy key without that capability cannot use MCP. Access covers the key’s workspace; the key’s project name does not limit results to that project. Use project filters to narrow a query.
The server runs on your infrastructure. Your choice of assistant and model determines where returned data is processed.
Read connection setup, tool arguments, and client examples →Investigate with read-only access
ShieldedStack’s MCP tools read inventory and findings. They do not change policy, edit repositories, acknowledge alerts, or reconfigure package managers. Your team can use the results to plan a change and carry it out through its normal workflow.
For findings that should reach people without an assistant query, configure Teams, Slack, PagerDuty, or webhook connections. MCP supports investigation; notification channels deliver alerts.
Explore related capabilities
Test it on your infrastructure
Use the 30-day on-premises trial to validate your package managers, policies, and operations. Prefer a first look? The separate 7-day hosted Portal evaluation uses synthetic package activity.