MCP Server

DriftGuard can run as a standalone MCP server, exposing its memory tools to any MCP-compatible client including Claude Desktop.

Start the server

driftguard-mcp

Available tools

Tool	Description
register_mistake	Store a causal memory in the mistake graph
register_success	Store a causal memory in the success graph
query_memory	Retrieve warnings and reinforcements for a context
deep_prune	Run a full cleanup pass over both memory graphs
graph_stats	Return node and edge counts for both graphs
guard_metrics	Return runtime review counters and gauges

Claude Desktop config

Add DriftGuard to your claude_desktop_config.json:

{
  "mcpServers": {
    "driftguard": {
      "command": "driftguard-mcp"
    }
  }
}

Tool usage examples

Register a mistake

register_mistake(
  action="increase salt",
  feedback="too salty",
  outcome="dish ruined"
)

Register a success

register_success(
  action="add more salt",
  feedback="well seasoned",
  outcome="dish praised"
)

Query memory

query_memory(context="add more salt")

Returns:

{
  "query": "add more salt",
  "warnings": [
    {
      "trigger": "increase salt",
      "risk": "too salty",
      "frequency": 3,
      "confidence": 0.87
    }
  ],
  "reinforcements": [
    {
      "trigger": "add more salt",
      "recommendation": "well seasoned",
      "frequency": 1,
      "confidence": 0.81
    }
  ],
  "confidence": 0.87
}

Get graph stats

graph_stats()

Returns:

{
  "mistakes": { "nodes": 9, "edges": 8 },
  "successes": { "nodes": 3, "edges": 2 }
}

Custom runtime config

Create a server with custom settings:

from driftguard import DriftGuardSettings
from driftguard.mcp import create_mcp_server

mcp = create_mcp_server(
    settings=DriftGuardSettings(
        storage_backend="sqlite",
        sqlite_filepath="driftguard.sqlite3",
        guard_policy="warn",
        retrieval_top_k=5,
    )
)

mcp.run()