Self-improving policy enforcement

The platformto enforce

AI agents ignore your rules. Your pricing logic, your refund policies, your escalation triggers, your compliance requirements. Open Bias catches violations before they reach your users.

View on GitHub
Works with your stack
Anthropic
OpenAI
Google Gemini
Meta Llama
Mistral
AWS Bedrock
LangChain
CrewAI
Anthropic
OpenAI
Google Gemini
Meta Llama
Mistral
AWS Bedrock
LangChain
CrewAI
Capabilities

Everything between your
agent and a bad decision.

01

Evaluator Engines

Four engines to match your enforcement needs. Judge uses a sidecar LLM for nuanced policy scoring. FSM handles deterministic workflows at zero latency. Choose the right tool for each policy.

02

Natural Language Policies

Write rules in plain English and compile them to executable policies. No DSL to learn, no config to debug. openbias compile "verify identity before refunds"

03

Automatic Intervention

Three strategies when policies are violated: amend system prompts, inject corrective messages, or modify responses directly. Enforcement happens before the next turn.

04

Observable by Default

Every enforcement decision is logged with per-condition pass/fail results. OpenTelemetry tracing, Langfuse integration, full audit trail. Answer "why did the agent do that?" with data.

00

Overview

Click any layer to expand.

01

WITHOUT GUARDRAILS

v

AI agents don't follow instructions reliably. They hallucinate, skip steps, and drift from policy — especially across long conversations and edge cases.

Identity, approvals bypassed · Policy divergence over turns · No audit trail · Every session, every user
02

MONITORING GAP

v

Your agents are running hundreds of sessions simultaneously. Some follow policy. Some don't. Without enforcement, you find out from a customer complaint — not a dashboard.

Hundreds concurrent · Unknown without enforcement · Customer complaints · None
03

PROXY

v

It sits between your app and the LLM provider. In between, every call is evaluated, logged, and — if needed — corrected. Zero changes to your agent code.

1 line change · 0ms (async default) · 4 pluggable · Open — never blocks prod
04

ENGINE

v

A refund flow needs deterministic enforcement. Tone of voice needs judgment. Open Bias lets you match the engine to the constraint, so you're not paying for LLM calls where regex would do.

0ms critical-path, async scoring · <1ms, deterministic, zero LLM cost · 100-500ms, handles ambiguity · Jailbreak, PII, toxicity
05

INTERVENE

v

Detection without action is just monitoring. Open Bias closes the loop: every violation triggers an intervention strategy that steers the agent back on course, automatically.

3 (prompt, inject, modify) · Deferred or immediate · Open — never blocks production · Full trace per intervention
Process

How it works

policy.yaml
1# policy.yaml
2policies:
3 - name: identity-check
4 trigger: "before refund"
5 conditions:
6 - "user identity verified"
7 action: amend_system_prompt
8 
9 - name: pricing-guard
10 trigger: "discount offered"
11 conditions:
12 - "discount <= 20%"
13 action: block_response
Ready
Metrics

Performance you
can measure.

 
0
Violations caught
0
Interventions triggered
<0ms
Added latency
0
Policies improved
Drift & Recovery

Deviation detection

Watch an agent drift from policy compliance in real time — and see the guardrails kick in.

Status
NOMINAL
PhaseSTART
Turn0 / 7
Confidence0.95
Session
Sessionses_8f3a
Workflowcustomer-support-fsm
Stategreeting
TierCONFIDENT
Enforcement Log
no interventions
POLICY0.02greeting
Integrations

Works with everything
you already use.

Connects to every major LLM provider, agent framework, and observability tool.

Anthropic
Anthropic
LLM Provider
OpenAI
OpenAI
LLM Provider
Google Gemini
Google Gemini
LLM Provider
Meta Llama
Meta Llama
LLM Provider
Mistral
Mistral
LLM Provider
AWS Bedrock
AWS Bedrock
Cloud Platform
Cohere
Cohere
LLM Provider
LangChain
LangChain
Agent Framework
CrewAI
CrewAI
Agent Framework
AutoGen
AutoGen
Agent Framework
OpenTelemetry
OpenTelemetry
Observability
Langfuse
Langfuse
Observability
NeMo Guardrails
NeMo Guardrails
Safety
Anthropic
Anthropic
LLM Provider
OpenAI
OpenAI
LLM Provider
Google Gemini
Google Gemini
LLM Provider
Meta Llama
Meta Llama
LLM Provider
Mistral
Mistral
LLM Provider
AWS Bedrock
AWS Bedrock
Cloud Platform
Cohere
Cohere
LLM Provider
LangChain
LangChain
Agent Framework
CrewAI
CrewAI
Agent Framework
AutoGen
AutoGen
Agent Framework
OpenTelemetry
OpenTelemetry
Observability
Langfuse
Langfuse
Observability
NeMo Guardrails
NeMo Guardrails
Safety
NeMo Guardrails
NeMo Guardrails
Safety
Langfuse
Langfuse
Observability
OpenTelemetry
OpenTelemetry
Observability
AutoGen
AutoGen
Agent Framework
CrewAI
CrewAI
Agent Framework
LangChain
LangChain
Agent Framework
Cohere
Cohere
LLM Provider
AWS Bedrock
AWS Bedrock
Cloud Platform
Mistral
Mistral
LLM Provider
Meta Llama
Meta Llama
LLM Provider
Google Gemini
Google Gemini
LLM Provider
OpenAI
OpenAI
LLM Provider
Anthropic
Anthropic
LLM Provider
NeMo Guardrails
NeMo Guardrails
Safety
Langfuse
Langfuse
Observability
OpenTelemetry
OpenTelemetry
Observability
AutoGen
AutoGen
Agent Framework
CrewAI
CrewAI
Agent Framework
LangChain
LangChain
Agent Framework
Cohere
Cohere
LLM Provider
AWS Bedrock
AWS Bedrock
Cloud Platform
Mistral
Mistral
LLM Provider
Meta Llama
Meta Llama
LLM Provider
Google Gemini
Google Gemini
LLM Provider
OpenAI
OpenAI
LLM Provider
Anthropic
Anthropic
LLM Provider
For developers

Ship policies,
not workarounds.

A proxy that gets out of your way. One line change to your client, full policy enforcement on every request.

Python native

Full type safety with Pydantic models.

YAML config

One file. Sensible defaults. No magic.

Provider agnostic

Anthropic, OpenAI, Google, Mistral, and more.

Open source

Self-host. No vendor lock-in.

pip install openbias
Use cases
01 / 04

"Enforce identity verification before processing refunds. Automatically amend system prompts when policies are violated."

C

Customer Support Agent

Use Case, Support Automation

Key Result

Policy violations caught

Pricing

Simple, scalable
pricing

MonthlyAnnualSave 17%
01

Open Source

Self-host the engines

Free
  • Unlimited agents & policies
  • OpenTelemetry tracing
  • Community support
Most Popular
02

Pro

Hosted enforcement with audit trails

Usage-based

Pay per policy evaluation. Coming soon.

  • Everything in Open Source
  • Managed cloud proxy
  • Policy dashboard & analytics
  • Audit trail with decision logs
  • Team management & RBAC
  • Priority support & SLA
03

Enterprise

Dedicated infrastructure for regulated environments

Custom
  • Everything in Pro
  • VPC / on-premise deployment
  • SSO & SAML
  • Compliance packages
  • Dedicated support engineer
  • Custom SLAs & contracts

Compare all features

What if your agents
couldn't break the rules?

Join teams enforcing AI agent policies in real-time. Open source. Free to start.

View on GitHub

No credit card required