Summary

Adds MaybeDontAnalyzer, a new SecurityAnalyzerBase implementation that validates agent actions against policy rules configured in a Maybe Don't Gateway instance.

• Calls the gateway's POST /api/v1/action/validate endpoint before action execution
• Maps ActionEvent fields (tool_name, tool_call.arguments, thought, summary) to the gateway's request format
• Maps the gateway's risk_level response directly to SecurityRisk (HIGH/MEDIUM/LOW/UNKNOWN)
• Graceful degradation: all errors (timeout, unreachable, 500, invalid JSON) return UNKNOWN
• Follows the GraySwan analyzer patterns: Pydantic fields, lazy httpx client, model_post_init env var resolution, set_events/close lifecycle

How it fits with existing integrations

MaybeDont doc for OpenHands

• https://maybedont.ai/docs/api/action-validate/
• https://maybedont.ai/docs/mcp-gateway/examples/connecting-openhands/

Configuration

from openhands.sdk.security.maybedont import MaybeDontAnalyzer

analyzer = MaybeDontAnalyzer()  # defaults to http://localhost:8080
# or
analyzer = MaybeDontAnalyzer(gateway_url="http://my-gateway:8080")
# or via env var
# MAYBE_DONT_GATEWAY_URL=http://my-gateway:8080

Files

Related

• docs: add Maybe Don't integration guides docs#350

Test plan

• uv run pytest tests/sdk/security/maybedont/ -v — 41 tests pass
• uv run pytest tests/sdk/security/ -v — 119 tests pass (no regressions)
• make format && make lint — clean
• Integration tested against a live Maybe Don't Gateway (CEL deny rule for rm -rf → HIGH, safe commands → LOW)

🤖 Generated with Claude Code