Real-time governance for AI agents. Use Orbitals.
Orbitals sits between users and AI systems, checking requests and responses against the principles your organization defines, then returning structured allow, block, or flag decisions with rationale.
Runtime trace
Every call gets an oversight decision.
User request
Can this assistant handle it?
Orbitals
Scope, policy, claims
AI agent
Only valid work flows through
Orbitals
Response checked before delivery
Response
Structured decision attached
Prompting is not runtime control.
AI agents eventually receive requests they should not answer: out-of-scope work, restricted actions, policy violations, or factually risky prompts. Without a runtime layer, the model will often respond anyway.
Prompts and generic filters cannot reliably enforce your product scope, policies, and brand standards before failures reach users.
Beyond today's AI oversight stack.
faster than frontier LLM-based oversight in internal single-GPU benchmarks
more cost-effective than frontier LLM oversight per million tokens in internal tests
fewer errors than frontier LLM oversight on safety and custom policy benchmarks
A lightweight oversight layer for every interaction.
Orbitals adds governance without fine-tuning the underlying model, rewriting your system prompt, or forcing a new platform. If your application already calls an LLM, integration is only a few lines of Python.
Inputs
Inspect user requests before they reach the model, so out-of-scope or restricted work can be blocked or routed early.
Outputs
Inspect AI responses before users see them, catching off-policy, off-brand, or factually risky content before delivery.
Decisions
Return allow, block, or flag decisions with structured rationale that engineering and compliance teams can read.
Custom SLMs, built for governance tasks.
Orbitals modules use purpose-trained multilingual small language models for low-latency governance tasks rather than general-purpose frontier models. Standard models are free on Hugging Face; Pro models are available through hosted or enterprise access.
Purpose-trained
Models built for governance tasks, not generic chat.
Low latency
Small language models designed for production request paths.
Multilingual
Designed for governance across multilingual AI systems.
Guard an AI service in a few lines.
Import the module, describe what your assistant is meant to do, validate the user request, and route violations before the model ever answers.
from orbitals.scope_guard import ScopeGuard
ai_service_description = "You are a helpful assistant for ..."
user_message = "Can I buy ..."
guardrail = ScopeGuard()
result = guardrail.validate(user_message, ai_service_description)
if result.scope_class.value in {"Restricted", "Out of Scope"}:
print("Request violates guardrail:", result.evidences)
else:
# The user request is safe.
# Pass it to the AI assistant for processing.Modular guardrails for high performance on the risks that matter to enterprise AI.
ScopeGuard
Scope and policy classification with conversation awareness, so agents stay inside the work they were designed to do.
ClaimExtractor
Intent and factual-claim extraction across user and assistant turns, ready for factuality and verification workflows.
More modules
Additional guardrails are coming for broader output policy checks, redaction, and attack-surface reduction.
Common questions about Orbitals AI guardrails.
Short answers for teams evaluating runtime governance, open-source Python guardrails, and small language models for LLM agents.
What is Orbitals?
Orbitals is an open-source Python library for real-time AI governance. It adds runtime guardrails around LLM agents by checking user requests and assistant responses before they move through the application.
How does Orbitals enforce AI guardrails?
Orbitals runs specialized governance modules on inputs and outputs, then returns structured allow, block, or flag decisions with rationale that engineering, product, and compliance teams can review.
Is Orbitals open source?
Yes. Orbitals is available as an open-source Python package, with standard governance models available for teams that want to start building AI guardrails without changing their underlying LLM stack.
Does Orbitals replace system prompts or moderation filters?
No. Orbitals complements prompts and moderation filters with a separate runtime oversight layer that can enforce product scope, custom policies, and factual-risk checks before failures reach users.
What are ScopeGuard and ClaimExtractor?
ScopeGuard classifies whether requests fit an AI service's scope and policy. ClaimExtractor extracts intent and factual claims so teams can build verification, factuality, and audit workflows around agent conversations.
Who is Orbitals for?
Orbitals is built for AI product, platform, safety, and compliance teams deploying LLM agents in workflows where requests, responses, policies, and audit trails need explicit runtime governance.