Skip to main content
LLM Gateway

Bring your model calls
under control.

The DVARA LLM Gateway connects your apps to model providers through one OpenAI-compatible API. As part of DVARA’s AI governance platform, it gives you a place to apply policies, protect sensitive data, and track usage.

Get startedBook a demo
Your appsApps & agents
DVARALLM Gateway
PolicyData protectionSpend controlsEnterprise
Configured request & response checks
FlightdeckUsage & auditEnterprise
OpenAI model providerAnthropic model providerGemini model provider
Model providersHosted or local
Live chat flow. Controls depend on your configuration and distribution; other API operations differ.

Set the rules for model access​

Give teams a shared place to manage model calls, without rebuilding the same controls in every app.

Decide which models apps can use

Define model restrictions in YAML policies and activate the rules you want enforced. Open Source evaluates Active policies loaded from files; Enterprise adds policy management in Flightdeck.

Explore policy controls →

Choose how sensitive data is handled

PII detection is on by default with the LOG action, which leaves content unchanged. Choose REDACT to remove detected values or BLOCK to reject a request containing them. Configure audit storage to retain detection events.

Review PII settings →

Set spending limits

In Enterprise, configured hard budgets reject new supported model calls once recorded spend reaches the limit. Accurate spend needs usage and model pricing. In-flight calls can still add cost; this is not a guaranteed final bill ceiling.

Explore cost controls →

Connect the models your teams use​

Use supported OpenAI SDK operations with DVARA as the base URL. Configure your providers and routes; check model and endpoint support before moving an app.

  • OpenAI
  • Anthropic
  • Gemini
  • Amazon Bedrock
  • Azure OpenAI
  • Mistral
  • Cohere
  • Groq
  • Ollama
  • Qwen
  • DeepSeek
  • Moonshot
  • ChatGLM
  • Grok

Streaming, tools, and structured outputs vary by provider. Check supported providers and the API reference.

Retry failed calls. Use a compatible fallback.​

Resilience and fallback are enabled by default. DVARA retries eligible failures and can try another registered provider that supports the same model request. It does not automatically translate a GPT model name into a Claude model.

Fallback candidates come from registered providers, not a route’s named fallback. If none succeeds, the request fails. Streaming fallback only covers opening the upstream stream, not errors while reading it.

Read the fallback rules →
Enterprise

See what happened in Flightdeck​

See recorded workspace usage and cost by model. With audit storage configured, use audit views to investigate recorded events and policy decisions. Cost reporting depends on available usage and model pricing.

Flightdeck records a denied mock/echo request for the Acme Alpha demo workspace, with the policy rule, trace ID and signing evidence expanded.Flightdeck records a denied mock/echo request for the Acme Alpha demo workspace, with the policy rule, trace ID and signing evidence expanded.
A recorded policy denial from a demo workspace. Colors adapt to this website’s theme; Flightdeck uses a light interface.

Explore Flightdeck · Understand audit evidence

Choose how to run it​

Open Source

A standalone LLM Gateway with file-based configuration, core policies, PII controls, and routing. Local audit is off until you configure its file path and signing secret. Run it in your own environment.

Start with Open Source →

Enterprise

Add Flightdeck, central configuration, budgets, and fleet operations. Manage model traffic alongside the platform’s MCP and A2A Gateways.

Explore deployment options →

See plans and production terms · Compare capability availability

For tool and agent calls, explore the DVARA MCP Gateway and DVARA A2A Gateway.

FAQ

Before you connect your app

Can I keep my OpenAI SDK?

Yes, for supported operations. Change the base URL to DVARA and use a DVARA API key. Check the API reference for your provider’s model names, request fields, and streaming support.

Can I use local models?

Yes. Connect models served by Ollama alongside supported hosted providers. Configure each connection and route; available features depend on the model and provider.

Does self-hosting keep every request local?

No. Self-hosting determines where DVARA runs. Calls to a hosted model still leave your environment to reach that provider; use a locally hosted model when requests must stay local.

Is sensitive data removed by default?

No. PII detection is on, but the default LOG action leaves content unchanged. Set REDACT to remove detected values or BLOCK to reject requests containing them. Audit retention needs configured storage; response and streaming controls have separate settings.

Will DVARA switch from GPT to Claude if OpenAI fails?

Not with the current built-in configuration. Retry is enabled by default for eligible failures, but fallback does not map one provider’s model to another. Do not rely on automatic cross-provider failover.

Start with one app.​

Connect a model, apply a policy, and check the result.
Want to talk it through? Bring your use case to a demo.

Get startedBook a demo