ARBR
Control Every AI Request
Self-hosted, MIT-licensed OpenAI-compatible control plane. Point existing SDKs at one gateway; ARBR classifies, routes, meters and governs each request before it hits a provider. Budgets, rate limits, output guardrails, prompt-injection checks and an application kill switch sit in the data path. Pinned models stay pinned; model auto follows only rules a human enabled. Runs standalone or above LiteLLM.
What is ARBR?
ARBR is a self-hosted, MIT-licensed OpenAI-compatible control plane. You point existing SDKs at one gateway; it classifies, routes, meters and governs each request before the provider sees it. It is for teams that need policy, budgets and reversible model changes across many providers, not just a proxy.
Key features
- Single OpenAI-compatible endpoint over major providers plus any OpenAI-compatible or self-hosted backend, including LiteLLM
- Per-request classify, route, meter and govern in the data path
- Policy before inference: budgets, rate limits, output guardrails, prompt-injection checks and an application kill switch
- Budget caps by application, provider, department, workflow, model or global, over a day or month, with alert, downgrade or block; a breached block cap returns 429 and outranks an explicit model pin
- Human-gated optimisation: observe traffic, surface model-switch opportunities, evaluate candidates, then shadow or canary with rollback; pinned models stay pinned
- Structured events for cost, latency, tokens and routing decisions, local dashboards or OpenTelemetry export
- Gateway API keys (prefix ab_) bind attribution and optional per-key RPM; keys shown once then stored as SHA-256
- MIT license, self-hosted; usable standalone or above LiteLLM
Who it's for
- A team running coding agents and apps across Anthropic, OpenAI, Gemini and a self-hosted endpoint who want one budget and one kill switch
- An engineering lead who will only move a workload to a cheaper model after shadow/canary evidence, not an automatic switch
- A company already on LiteLLM that needs governance, evals and measured rollouts above the proxy
- Anyone who must attribute spend by application, team and user and export that to Datadog, Grafana or Prometheus
When not to use it
Not a drop-in replacement for a simple multi-provider proxy: if you only need to fan out API keys, LiteLLM or a managed gateway is enough. Also not a hosted SaaS you can try without deploying it — the product is self-hosted, and the site's hosted option is not priced on the public page.
FAQ
Is ARBR free?
The control plane is open source under the MIT License and self-hosted. Product Hunt lists the launch as Free. A hosted option is linked from the site but public hosted pricing is not itemized.
Does it override a model I pinned in code?
No. The site and README state that an explicitly pinned model is honored. When an app sends model auto, ARBR follows only the rules and policies a human has enabled.
Can I keep LiteLLM?
Yes. ARBR can sit above LiteLLM for analysis, recommendations, evaluation gates and measured optimisation, or run as a standalone gateway.
What does a budget cap actually do?
Docs: each cap has a dimension, period, dollar limit and action (alert, downgrade or block). Downgrade and block are enforced before routing. A breached block cap returns HTTP 429 and outranks even an explicit model pin for gateway-enforced dimensions.
Share this launch
Embed this badge
<a href="https://orangebot.ai/product/arbr" target="_blank" rel="noopener noreferrer"> <img src="https://orangebot.ai/api/badge/arbr.svg" alt="Featured on OrangeBot" width="200" height="54" /> </a>