CLI Reference

driftguard-mcp

Start the DriftGuard MCP server:

driftguard-mcp

Uses DriftGuardSettings defaults — JSON persistence, warn policy, standard embedding model.

To customize, create a server.py entry point:

from driftguard import DriftGuardSettings
from driftguard.mcp import create_mcp_server

mcp = create_mcp_server(
    settings=DriftGuardSettings(
        storage_backend="sqlite",
        sqlite_filepath="driftguard.sqlite3",
    )
)

mcp.run()

driftguard-benchmark

Run the built-in quality benchmark suite:

driftguard-benchmark

Output formats

Human-readable text (default):

driftguard-benchmark --format text

Machine-readable JSON:

driftguard-benchmark --format json

Flags

Flag	Default	Description
--format	text	Output format: text or json
--log-level	WARNING	Package log level during the run

Sample output

DriftGuard Benchmark Report

Merge: precision=1.00 recall=0.50 f1=0.67
- seasoning paraphrase: PASS (expected='increase salt', predicted='increase salt')
- heat paraphrase: FAIL (expected='raise heat heat', predicted=None)
- unrelated action: PASS (expected=None, predicted=None)

Retrieval: precision=1.00 recall=0.75 f1=0.86
- seasoning warning: PASS
- heat warning: PASS
- resting warning: PASS
- unrelated no-warning: PASS

Demo agents

Two local demo agents are included in the repository:

Rule-based demo

python demo/rule_based/demo_agent.py \
  --duration-seconds 120 \
  --step-delay 2 \
  --prune-every 4 \
  --reset-graph

LangGraph LLM demo

export OPENAI_API_KEY=your-key
python demo/langgraph/demo_agent.py \
  --duration-seconds 120 \
  --step-delay 4 \
  --prune-every 4 \
  --reset-graph

Demo flags

Flag	Description
--duration-seconds	How long to run the loop
--step-delay	Seconds between steps
--prune-every	Run deep prune every N steps (0 to disable)
--reset-graph	Delete graph and trace before starting
--runtime-mode	demo (offline) or real (full embedding stack)
--log-level	Package log level