Is what it says grounded?
Response-level checks on whether the model's output is true and supported — the same family of signals Director-AI provides.
Content · integrity · action
Govern what an agent does — not just what it says.
Director-Class AI sits between an autonomous agent and its effectors. It reviews high-impact shell, SQL, infrastructure, API, and MCP actions before they dispatch; uncertain or high-risk actions escalate to human approval, and every decision emits a tamper-evident audit record. A separate commercial product built above the Apache-licensed Director-AI base.
Effector-boundgoverns the action, before dispatch
Human-in-loophigh-risk escalates for approval
Tamper-evidentaudit & evidence on every verdict
CommercialBUSL-1.1, source-available
One verdict, three planes
Most guardrails check the words. This one checks the act.
A parallel detector ensemble governs three planes at once. The action plane is the differentiator; content and integrity are supporting signals that feed effector-bound governance, not a standalone prevention guarantee.
Content
Integrity
Was the input tampered with?
Detects manipulation of the input or surrounding context before it shapes a decision.
Action — the differentiator
Is what it does safe to run?
The shell, SQL, infrastructure, API, and MCP call the agent is about to make is reviewed before dispatch; destructive or high-risk actions are blocked or escalated.
What it does
A control point between the agent and the real world.
Pre-dispatch action review
High-impact shell, SQL, infrastructure, API, and MCP tool calls are evaluated before they execute — not after the damage is done.
Human escalation
Uncertain or high-risk actions are held and routed to a human for explicit approval instead of dispatching automatically.
Tamper-evident evidence
Every verdict emits an audit and evidence record, so an agent's real-world actions carry a durable, reviewable trail.
SDK, MCP gateway, SIEM export
Drop it in as a library, run it as an MCP gateway in front of your tools, and stream decisions to your existing SIEM.
The action checkpoint
Five lines between intent and effect.
# gate a destructive command before it runs from director_class_ai.action import DestructiveCommandDetector from director_class_ai.core import ParallelEnsembleScorer, EvaluationRequest guard = ParallelEnsembleScorer([DestructiveCommandDetector()]) verdict = guard.evaluate(EvaluationRequest(action="rm -rf /")) assert verdict.allow is False assert verdict.requires_human is True
The same checkpoint wraps any effector-bound action your agent takes — a shell command, a SQL statement, an infrastructure API call, an MCP tool invocation.
A blocked action does not silently fail: it returns a verdict with the reason, the requires-human flag, and the evidence record that explains the decision.
Built above the Apache-licensed Director-AI base, which it reuses for the content and integrity planes.
Honest status
Production-ready. Commercial access on request.
Director-Class AI is production-ready and source-available under the Business Source License 1.1. Commercial pricing is being finalised — reach out for early commercial access and an evaluation. Benchmark evidence to date is local functional evidence; we do not position it as a public advantage claim.