DVARA 1.0.0-GA — Governance for Every LLM and MCP Call
DVARA 1.0.0 is generally available today. We are launching an AI governance platform — the LLM Gateway and the MCP Proxy as components — for engineering teams running real production traffic and dealing with real auditors, real compliance officers, and real cost lines on a real budget.
This post covers why we built it, what is in 1.0.0-GA, and how to run it in your own infrastructure today.
This post originally said "there is no feature paywall." That was accurate for 1.0.0 — DVARA was a single edition at the time, and any valid license unlocked every feature. DVARA 1.4.0 introduced a free Community Edition, so there are now two editions. The MCP Proxy, signed audit, Policy-as-Code, PII redaction, and cost attribution described below are Enterprise-only.
Why DVARA exists
If you have been routing LLM traffic through a Python proxy on a single VM, you already know the failure mode that pushed us to build this. Things work. Things keep working. Then production catches up.
Production catches up the day a compliance reviewer asks for proof that no PII left your perimeter, and the answer is a stack of unsigned text files. It catches up the day finance asks who spent the $14K on Anthropic last week and the answer is one shared API key. It catches up the day your first agent ships, hits a tool-call loop in the middle of the night, and burns through a budget cap that does not exist yet.
The existing options solve part of the problem. LiteLLM is excellent for prototyping — a single binary that hides the differences between OpenAI, Anthropic, and a long tail of providers. We respect what that project did for the ecosystem; this post is not an attempt to argue otherwise. The constraint is that LiteLLM was built for a developer routing their own calls, not for a platform team governing many teams at once. Multi-tenant cost attribution, signed and chained audit trails, Policy-as-Code that a compliance reviewer can read, and tool-call governance for the agents your company is about to ship — those problems land somewhere else.
Portkey is a managed SaaS with a unified API, observability, guardrails, and an MCP gateway. It is more polished than rolling your own — and its paid plans meter on log volume and retention, so the bill climbs as your traffic and history grow. For a team trying to forecast cloud spend twelve months out, that variable line is a planning problem before it is a billing problem. SaaS shape, SaaS uncertainty.
DVARA is what you deploy when the workload is past prototyping and the people asking questions about it are no longer all engineers. Drop-in OpenAI compatibility is still there — we kept the easiest part — but the centre of the product is governance. Every call is policy-evaluated, PII-scanned, audited, and cost-attributed before it leaves your perimeter, on call one. Routing, failover, structured outputs, and semantic cache exist in the same product, but they exist to serve the governance story, not the other way round.
What is in 1.0.0-GA
DVARA 1.0.0 is self-managed: you run it in your own infrastructure — a Docker image or a Helm chart — activated by a signed license key, with bring-your-own-key credentials so provider keys and request data never leave your perimeter. Start with a 30-day, all-features trial license (every feature unlocked, no credit card), then license annually for production. Every Enterprise install ships the full governance core, and Agentic AI and the DVARA MCP Proxy (the governed tool-call proxy, human approval gates enforced at execution, agent loop detection with auto-kill, and the unified LLM + MCP session timeline) are included in every Enterprise install alongside signed audit, Policy-as-Code, PII detection, and request-level cost attribution.
The platform supports OpenAI, Anthropic, Google Gemini, AWS Bedrock, Azure OpenAI, Ollama, Mistral, Cohere, Groq, Qwen, DeepSeek, Moonshot Kimi, ChatGLM, and xAI Grok on Day 1, all through a drop-in OpenAI-compatible interface. Capability-aware route filtering means a request asking for json_schema does not get routed to a provider that cannot natively support it. Failover, structured outputs, streaming, and a per-route circuit breaker are along for the ride.
Governance is the centre. Every event in the audit trail is cryptographically signed and chained at write time, so any tampering is detectable end-to-end. The Policy-as-Code engine takes a YAML DSL that a compliance reviewer can read, supports dry-run against historical traffic, and can be promoted from SHADOW to ACTIVE without restart. PII detection scans inputs and outputs and supports BLOCK / REDACT / LOG actions per tenant. Cost attribution lives at the request level, not the daily aggregate, so the spend for one tenant on one model on one day is a query, not a forensic exercise. SOC 2, HIPAA, and GDPR evidence packages are generated on demand from the audit trail.
If you would rather not operate DVARA yourself, a fully managed option — we run a dedicated cluster for you, in your region — is available alongside the self-managed install. Production licensing is annual and sized to your team; talk to us for a plan, or for founder-led help getting DVARA into production.
DVARA vs Portkey
If you are on Portkey today, watching the metered line items climb with your log volume and retention, the contrast is the billing shape and where your data lives. Portkey is a managed SaaS that meters on usage; DVARA runs in your own infrastructure under an annual license — your traffic and audit history stay in your perimeter, and the cost is one predictable line, not a variable that grows with retention.
The migration is two lines of code, because the drop-in OpenAI compatibility is the same. Point your OpenAI SDK's base_url at your own DVARA gateway, add your provider keys in the DVARA Flightdeck, and existing application code keeps running — streaming, structured outputs, tool calls, all of it.
The bigger differentiator is governance depth: signed and chained audit, per-tenant Policy-as-Code a reviewer can actually read, and PII detection that enforces, not just logs. And when your roadmap reaches agents, every DVARA install includes the MCP Proxy — argument-level tool policies, human approval gates enforced at execution, and the same signed audit extended to every tool call.
We wrote a longer migration walkthrough at /migrate-from-portkey. Same shape from LiteLLM Cloud at /migrate-from-litellm-cloud.
To see DVARA on your own traffic, get a 30-day trial license — every feature unlocked, including the MCP Proxy and agentic governance, no credit card. For production licensing or a managed deployment in your region, talk to us. The launch is live now.