Director-Lite
A three-line guard facade with a model-free heuristic default and a facts/RAG handoff. Optional NLI upgrade when you need it.
Free · pip install director-ai-lite
Verify the answer · gate the action
Catch the hallucination before it ships.
Director-AI scores every answer against retrieved evidence through five tiers — rules, embeddings, NLI — and halts unsupported output at the token level. The commercial edition extends the same gate to an agent's real-world actions.
14.6 msverify latency, NLI tier
75.8%balanced accuracy, LLM-AggreFact
<0.5 msheuristic tier
FreeLite tier on PyPI
Try the gate
Verify a claim, or gate a risky action. Watch it pass or halt.
An illustrative, client-side version of the gate logic — production uses real NLI scoring and policy rules. Nothing you type leaves your browser.
Try editing the claim — drop “audited by KPMG” and it passes.
Three editions
Start free. Grow into output, then action control.
One guardrail, three commitments — drop in a three-line guard, run the full open core, or gate an autonomous agent's real-world actions.
Director-AI
The open-core runtime: 5-tier scoring, REST/gRPC server, SDK and framework integrations, evidence packets, and the opt-in token-level halt.
Free core · commercial licence to ship closed-source
Compare licences →Director-Class-AI
Runtime action control. Reviews high-impact shell, SQL, infrastructure, API, and MCP actions before dispatch; routes high-risk ones to human approval — then pairs with HushLine to contain the execution: secrets redacted from output, directories gated.
Commercial engagement · custom
Talk to us →Quickstart
Three lines between your model and a hallucination.
pip install director-ai
from director_ai import guard
verdict = guard(answer, evidence)
if not verdict.supported:
answer = verdict.safe_fallback # halt the unsupported claimguard() returns a verdict with the score, the tier that decided it, and the evidence behind it — so every decision is auditable.
Point it at Remanentia recall to check answers against your own indexed sources, not just the prompt.
All integration surfacesWorks with any modelOpenAI, Anthropic, or a local model — the guard sits beside your model, it doesn't replace it.
Runs where your agent runsSDK, middleware, or sidecar. Self-hosted and local-first by default, no mandatory SaaS.
Your data stays localEvidence and prompts never have to leave your network to be checked.
How the gate works
Five tiers, cheapest first. Stop as soon as you can.
Each answer climbs only as far as it needs to. Rules and heuristics settle the easy cases for free; embeddings and NLI handle the rest. The streaming halt cuts an unsupported claim mid-token instead of after it ships.
1Rules & format checks reject malformed or policy-violating output for free.
2Heuristics score claims against compiled facts without a model call.
3Embeddings measure semantic support from the retrieved evidence.
4NLI decides entailment vs contradiction on the claims that are still uncertain.
5The streaming halt stops an unsupported answer at the token level, before it leaves.
No marketing math
Most guardrails sell a “99% reliability” number. We publish the command.
Director-AI's headline is a named public benchmark — balanced accuracy on LLM-AggreFact — not a round figure with no method behind it. Every number ships with the script that produced it, and the streaming halt is labelled opt-in, not sold as a silent guarantee.
✓A named public benchmark (LLM-AggreFact) — not a private “reliability” score.
✓A reproducibility command beside every number.
✓The opt-in halt is marked opt-in, not a guarantee.
✓Latency you can verify: 14.6 ms NLI, <0.5 ms heuristic.
How it compares
Where Director-AI is different.
Generalised from the public positioning of hallucination-guardrail vendors as of June 2026. Specifics vary by plan.
| Director-AI | Typical vendor | |
|---|---|---|
| Verify latency | 14.6 ms NLI · <0.5 ms heuristic | 50–200 ms inline |
| Headline metric | Named public benchmark + repro command | Private “reliability %” |
| Pricing | Public, one-click checkout | “Contact sales” |
| Deployment | Local-first, self-hosted | SaaS-first |
| Source | Open core on PyPI | Closed |
| Entry | Free (Lite) or CHF 290/yr | Demo-gated |
What ships in the core
A guardrail you can run anywhere your agent runs.
guard() SDK
One call wraps any generation path and returns a verdict with the supporting evidence.
REST & gRPC server
Run the guard as a sidecar service; FastAPI middleware drops it into existing apps.
Evidence packets
Every verdict carries the sources and scores behind it, so a decision is auditable.
Injection detection
Catches prompt-injection attempts in tool output and retrieved content.
Agent & MCP preflight
Checks an agent's planned tool call before it runs, not after.
Rust accelerators
12 Rust-accelerated compute paths keep the hot scoring loops fast.
Ready to ship