One thin MCP server per security platform. Each server imports
core and contains only what is platform-specific: the API client
and the tool definitions.
Pattern (see CONTRIBUTING.md and ../CLAUDE.md):
- Read tools first; target ≤ ~8 flat tools per server.
- All output goes through
core(schema, redaction, paging) — never hand-rolled. - Write actions route through
core/gating/(config flag + confirmation token). - Ship
.env.<platform>.example, contract tests, and anevals/task set.
Built & live-validated (8): defender-mcp, entra-mcp, limacharlie-mcp,
projectachilles-mcp, projectachilles-actions-mcp (gated writes),
intune-mcp, tenable-mcp, purview-mcp. Each server's README documents its
tools, required credentials/permissions, and smoke test.
Planned: wazuh-mcp, elastic-mcp, splunk-mcp, sentinel-mcp,
crowdstrike-mcp, sentinelone-mcp, sophos-mcp, misp-mcp, thehive-mcp,
opencti-mcp.