guard() SDK
Wrap any generation call and get back a verdict plus the supporting evidence.
from director_ai import guard
verdict = guard(answer, evidence)Integrations
One call wraps any generation path.
Drop the guard in as a function, a middleware, or a sidecar service — then point it at whichever model and framework you already run.
Deployment surfaces
Four ways to run the guard.
FastAPI middleware
Add the guard to an existing API with a single middleware registration.
app.add_middleware(DirectorGuard)REST & gRPC server
Run the guard as a sidecar; any language calls it over HTTP or gRPC.
director-ai serve --port 8088Agent & MCP preflight
Check an agent's planned tool call before it runs, over the MCP protocol.
director-ai mcp-preflightFrameworks & models
11 connectors, model-agnostic.
The guard sits beside your stack — it does not replace your model or your framework.
OpenAI
Anthropic
LangChain
LlamaIndex
CrewAI
Haystack
Semantic Kernel
FastAPI
REST
gRPC
MCP
Inside the runtime
Wired to memory and the agent channel.
Evidence from Remanentia
Point the guard at Remanentia's recall so verdicts check against your own indexed sources, not just the prompt.
Hand-offs over Synapse
A halted answer or a gated action can be routed to another agent or to a human over the Synapse channel.
MCP across the suite
Memory, verification, and coordination all speak MCP, so they compose without bespoke glue.