Spend nobody can explain
The invoice arrives as one number per provider. Finance cannot tell which app, team or workflow drove it, so there is nothing to optimise.
Open source by Gyde · MIT licensed
Scale AI across the enterprise. Keep control of every request.
ARBR is Gyde's open-source AI gateway. It sits between your applications and every model provider, so you can route each request to the right model, see who is spending what, and enforce limits before a call ever leaves your stack.
One OpenAI-compatible endpoint. Self-host it or use the hosted version.
01
Identify
Which app, team and user sent this, and what kind of task is it?
02
Check policy
Is it inside budget, rate limits and guardrails?
03
Pick the model
Rules, cost and task difficulty decide. Fallback if a provider fails.
04
Record it
Cost, latency, tokens and the routing reason, logged as one event.
Routed to the model your rules chose
Attributed · Logged · Within budget
One control layer across the providers your teams already use
OpenAIAnthropicGoogle GeminiAmazon BedrockAzure OpenAIVertex AIGroqDeepSeekMistralSelf-hosted models
The problem after the pilot
The first AI use case is easy. The twentieth is where control breaks.
Once AI moves past a single team, every group picks its own provider, keys and habits. Nobody planned the sprawl, but everyone inherits it.
The most expensive model, by default
Teams reach for the model they tried first. Simple summaries and extractions run on frontier models that the task never needed.
Policies written down, not enforced
Usage policies live in a document. Budgets, rate limits and prompt-injection checks are not applied at the moment a request is made.
One provider, one point of failure
When an app is hard-wired to a single provider, an outage, price change or deprecation becomes an engineering project instead of a config change.
About Gyde
Built by the team that takes enterprise AI from pilot to production.
Gyde is an enterprise AI company. We build full-stack AI systems for large organisations across financial services, manufacturing, retail, telecom and healthcare.
Getting people to use AI is one half of the job. The other half is running it responsibly once they do: knowing which models are in use, what they cost, and whether every request follows policy.
ARBR is the control layer we built for that second half. We released it as open source so any team can inspect it, run it and extend it, with or without working with Gyde.
Your people
Employees, teams and leaders
Where Gyde runs AI adoption, training and enablement.
Your applications
Copilots, agents, internal tools
Built in-house, by partners, or with Gyde.
The control layer
ARBR
Routing, budgets, guardrails, evaluation and a full record of every request.
Your providers
Any model, any vendor
Commercial APIs, cloud platforms or models you host yourself.
AI Services
Model Routing
Design routing policies and model mix for your workloads, with ARBR or your existing gateway.
Explore Model Routing →︎ Strategy and GovernanceEnterprise AI Governance
Turn AI policy into controls that are enforced in systems, not just written in documents.
Read the governance guide →︎ Adoption and EnablementAI Adoption Programmes
Help employees, departments and leaders put AI to work in their daily jobs.
See AI adoption →︎How it works
One gateway. Every decision made before the model runs.
Applications send requests to ARBR instead of directly to a provider. ARBR handles the rest in the request path.
Connect
Add your provider keys once. Applications point at a single OpenAI-compatible endpoint.
Route
Rules, cost guardrails and difficulty-aware selection choose a model per request, with automatic fallback.
Govern
Budgets, rate limits, output guardrails, prompt-injection checks and kill switches act before inference.
Improve
Score live traffic with an LLM judge, canary or shadow new models, and roll back if quality drops.
Built for every owner of AI
Each team gets the answer it has been asking for.
ARBR turns every model call into a structured event, so engineering, finance and security work from the same record.
Ship faster without a provider rewrite
Keep the SDKs and frameworks you already use. Change models through configuration, not code.
- Drop-in OpenAI-compatible API
- JavaScript and Python SDKs
- Streaming, embeddings and realtime voice proxy
Know where every rupee and dollar goes
Attribute spend by application, team, model and user. Set budgets that stop overruns instead of reporting them later.
- Cost per app, team and workflow
- Hard budgets and rate limits
- Savings recommendations from real traffic
Enforce policy at the point of use
Apply guardrails consistently across every provider, with an audit trail of what was sent, where and why.
- Prompt-injection and output guardrails
- Kill switches per app or model
- Self-host inside your own infrastructure
Prove quality, not just lower cost
Measure output quality on live traffic, then move each task to the cheapest model that clears your bar.
- LLM-judge evaluation on sampled traffic
- Canary and shadow deployments
- Regression gates with instant rollback
Integrate in minutes
Change two lines. Keep your stack.
Point your existing OpenAI client at ARBR. From the next request, you get routing, logging, authentication and budget enforcement.
- Works with existing OpenAI SDK code
- Compatible with agent frameworks and chat UIs
- Export telemetry to OpenTelemetry tools
# Same OpenAI SDK. New base URL. from openai import OpenAI client = OpenAI( base_url="https://arbr.yourcompany.com/v1", api_key="ab_your_gateway_key" ) response = client.chat.completions.create( model="auto", # let ARBR choose messages=[{"role": "user", "content": "Classify this support ticket"}] )
Choose how you run it
Open source when you want control. Hosted when you want speed.
Same control plane either way. Start hosted, move in-house later, or the other way round.
ARBR, managed for you
Sign up and start routing requests without running any infrastructure.
- No servers to operate
- Fastest path to a first governed request
- Good fit for teams evaluating ARBR
ARBR, inside your infrastructure
Run the full control plane in your own environment, with your data staying where you put it.
- Full source code, no lock-in
- Fits strict data-residency requirements
- Extend and contribute on GitHub
FAQ
Questions teams ask before they switch.
What is ARBR?
ARBR is an open-source, MIT-licensed AI gateway and control plane built by Gyde. It sits between your applications and AI model providers, and routes, meters and governs every request through one OpenAI-compatible endpoint.
Do we need to rewrite our applications to use ARBR?
No. ARBR is OpenAI-compatible. Teams change the base URL and API key in their existing OpenAI SDK, agent framework or chat UI, and requests start flowing through ARBR.
Can ARBR be self-hosted?
Yes. ARBR is open source and can run inside your own infrastructure. A hosted version is also available for teams that do not want to operate it themselves.
Which model providers does ARBR support?
ARBR works across major providers including OpenAI, Anthropic, Google Gemini, Amazon Bedrock, Azure OpenAI, Vertex AI, Groq, DeepSeek, Mistral and xAI, plus any OpenAI-compatible or self-hosted endpoint.
How is ARBR different from calling a model provider directly?
Calling a provider directly gives you one model and that provider's console. ARBR adds a neutral layer across all providers, so routing, budgets, guardrails, fallback and cost attribution are applied consistently, regardless of which model serves the request.
How does ARBR relate to Gyde?
Gyde is an enterprise AI company that runs AI adoption and enablement programmes and builds full-stack AI systems for large organisations. ARBR is the open-source AI gateway Gyde built and maintains, so the AI systems an enterprise runs stay observable, governed and cost-efficient as usage grows.
Put one control layer in front of every model.
Route your first request through ARBR today and start building a governed view of your AI usage.