Core concepts
The mental model behind watches, drift classification, and coverage enforcement.
Watches
A watch is a monitored external dependency — one URL or MCP endpoint on a schedule. Learn more →
Change records
When a snapshot differs from the previous baseline, DriftGuard creates a change record (also called a drift event) with classified changes. Learn more →
MCP monitoring
MCP watches track tools/list shape — tool removal is always breaking for agents. Learn more →
Coverage assert
CI scans your repo for dependency URLs and fails if any are unwatched. Learn more →
Glossary
Contract drift vocabulary — watch, snapshot, drift_status, pre-run check, agentAction. Glossary →
Five-layer map
Detection → status → policy → remediation → agent MCP surface. Five-layer map →
Severity rules
- breaking — requires code or agent changes before safe deploy
- warning — review recommended; may break strict clients
- info — backward compatible; documentation-only