Skip to main content
Sage’s graph analysis is backed by the BloodHound MCP server, which Sage runs as a subprocess and connects to at startup. BloodHound is a pre-wired, dedicated MCP server with its own agent; to attach any other MCP server, see Connecting MCP servers.

​
Point Sage at the MCP server

Set the MCP checkout directory before starting Sage. The container image bakes its own default (ENV SAGE_BLOODHOUND_MCP_DIR=/opt/bloodhound_mcp), so under Mythic you can leave this commented unless you are pointing at your own checkout. For local development, uncomment in Payload_Type/sage/.env:

​
Provide credentials

The MCP server needs to reach your BloodHound CE instance. Credentials resolve through the standard load order, so there are two places to set them without a shell:
  1. Mythic chat configuration when you create the chat (per-chat, the same place your model API key lives). Mythic user secrets work too and keep the token out of plaintext config.
  2. Sage’s own .env (InstalledServices/sage/.env under Mythic), edited from the Mythic web UI and shared by every chat in that container.

​
The connection is process-global: configure once

The first chat that connects successfully establishes the BloodHound connection for the whole container; every later chat reuses it. Fill these fields in once and leave them blank in subsequent chats.
  • A failed connect establishes nothing; the next chat tries again from scratch. You are not stuck with a bad connection.
  • Plain /bloodhound is idempotent: it reports an existing connection and is a safe way to ask whether BloodHound is connected.
  • /bloodhound force rebinds an already-connected container to the current chat’s resolved credentials without a restart. The rebind disconnects before connecting, so if the new settings are wrong you are left with no connection rather than the old one. /bloodhound will connect again once you have fixed the configuration.

​
When BloodHound is not connected

Sage still works. A missing optional dependency degrades a capability, not the product:
  • Ordinary conversation and supervised chats answer normally with no BloodHound credentials.
  • Autonomous solves fail closed, on purpose. A solve reasons over the attack graph, so running one without the graph would mean acting blind. The request is refused with a message that names BloodHound and repeats the setup steps.
  • Nothing is retried pointlessly: when required credentials resolve nowhere, Sage does not spawn an MCP server it knows will exit. Fix the configuration and the next request tries again.
In the Sage container log you will see, at WARNING:
It names which credentials arrived and which did not, never their values. A bare McpError: Connection closed with no explanation means you are on a build from before this was fixed.