The AI Governance
Control Plane
Arbiter sits between your applications and your AI providers. Every request passes through it. Budgets, policies, and guardrails activate automatically.
Works with Anthropic and OpenAI. One URL change. No other code modifications required.
Before
base_url = "https://api.anthropic.com" headers = { "Authorization": "Bearer sk-ant-..." }
After — Full Governance Active
base_url = "https://api.arbiter.ariva.ai/v1/anthropic" headers = { "Authorization": "Bearer sk-ant-..." "x-arbiter-api-key": "cg-..." }
Governance Modules
Six Layers of Control
Each module has four operating modes: off / observe / warn / enforce. Start in observe, move to enforce when ready. A module can never block unless you configure it to.
Observability
Every request logged from the first call. Provider, model, token counts, cost, latency, and attribution — captured automatically. No configuration required.
- Per-request cost tracking in real time
- Attribution by user, project, and environment
- Paginated request log with full filter controls
Budget Controls
Set spending limits at the workspace or per-user level. Requests that would breach the limit are blocked before the model is ever called.
- Monthly workspace and per-user spend limits
- Alert thresholds at configurable percentages
- Observe, warn, or enforce modes
Enforcement
Block requests before they reach the AI provider based on deterministic rules — model, team, user, or spend limit.
- Block by model, team, user, or spend threshold
- Rules evaluated oldest-first; first match wins
- Full audit log of every blocked request
Model Routing
Transparently redirect AI requests from one model to another. The application sends a request for Model A and receives a response from Model B — with no code changes.
- Cost optimisation — route low-complexity requests to cheaper models
- Fallback — redirect when primary model is unavailable
- Policy enforcement across teams without deployment changes
Usage Policies
Monitor post-request patterns and surface alerts when thresholds are crossed. Detects abnormal usage before it becomes an incident.
- High-cost request detection
- Premium model overuse alerts
- Repeated expensive prompt pattern detection
Output Guardrails
Detect and suppress dangerous AI-generated content before it reaches your application or end users. Shell commands, SQL mutations, credential leaks, infrastructure actions.
- Pre-delivery interception — blocks before the client receives the response
- Shell, SQL, infrastructure, secrets, Git, Kubernetes categories
- Observe baseline → warn → enforce progression
Positioning
Visibility Is Not Governance
Helicone and Portkey give you a cost dashboard. Arbiter gives you a governance control plane with enforcement, access control, audit logs, and output safety — without stitching three products together.
| Capability | Helicone | Portkey | Kong | Arbiter |
|---|---|---|---|---|
| Observability | ✓ | ✓ | ✓ | ✓ |
| Cost tracking | ✓ | ✓ | ✓ | ✓ |
| Model routing | — | ✓ | ✓ | ✓ |
| Budget enforcement | — | Partial | Partial | ✓ |
| User / team blocking | — | — | Partial | ✓ |
| Output guardrails | — | Via integrations | Via plugins | Native |
| Audit logs | — | Partial | ✓ | ✓ |
| Enterprise RBAC | — | Partial | ✓ | ✓ |
| Consulting-led | — | — | Sometimes | ✓ |
Who Arbiter Is For
Developer tooling
"I want a cost dashboard and usage charts."
Helicone, Portkey, and Braintrust do this well. They are built for developers who want observability.
Not what Arbiter is for.
Enterprise governance
"I need to control what AI can do, who can use it, and prove it."
Arbiter is for the CTO or CISO who needs enforcement: spending limits that actually block, models that can be restricted by team, dangerous outputs that are intercepted, and an audit trail that satisfies compliance.
Different buyer. Different value proposition. If you need a governance control plane — not just a dashboard — Arbiter is built for that.
Pilot Access
Arbiter is currently available to a limited number of pilot customers. Pilot engagements are consulting-led and include onboarding support.
Request access through a private governance review. We will assess fit and provision your workspace if aligned.
Request Pilot AccessWorks with Anthropic (Claude) and OpenAI (GPT-4, o-series).