Product
One appliance, five connected stages, on your infrastructure
MERIDVAR combines an enterprise AI governance layer with an admission gateway that can refuse an AI request before the model provider is contacted. You install it and you operate it. There is no MERIDVAR cloud.
- Software you install and operate
- Runs on a Docker host on your private network
- One appliance for one organisation
- Nothing is sent to MERIDVAR
Gateway decision Refused
A request carrying a national identification number, sent from the console playground.
- Provider calls
- 0
- Stored sample
- Masked
The refusal names the rule that caused it. Nothing was sent, and nothing is billed.
The path of a request
What enters, what is checked, and what is left behind
What enters
A request arrives through the OpenAI-compatible API, the console playground or the proxy path.
What is analysed
The whole request is read: string content, content blocks, system messages, tool call arguments and tool results.
Where policy applies
A pipeline of seventeen stages applies your policy, from identity and shadow AI to prompt DLP, data residency, budgets and approved models.
What is blocked
Any licensed, enforcing stage may refuse. Every later stage is skipped, the provider is never called and nothing is billed.
What is allowed
An admitted request is sent to Anthropic, OpenAI, Azure OpenAI, xAI or a self-hosted model endpoint.
What is recorded
Each decision is written to an append-only, hash-chained log, ready for an audit package. Findings keep masked samples.
The five stages
Discover, assess, approve, control and prove
Every capability below is documented as shipped in the current release.
01 Discover
What AI is in use?
- AI inventory. A register of AI systems observed in traffic or declared by a person, each with its provenance, owner, purpose, vendor, models and lifecycle status.
- Shadow AI discovery. Traffic compared with your list of sanctioned providers, in off, observe or enforce mode. Proxy and DNS logs can be uploaded for traffic that never touched the gateway.
- Agents and MCP servers. An agent registry with signed identities, and a declared MCP server inventory with tool-level risk. Endpoint agents report local AI tools.
AI inventory
Systems that this appliance observed in traffic or that a person declared.
An observation never becomes a declaration on its own. A person confirms it.
02 Assess
How risky is it?
- Versioned questionnaires. Three questionnaires ship with the product: an AI system baseline, AI vendor due diligence, and an AI agent and MCP server review. A published version never changes.
- Rules-based risk classification. Low, Moderate, High or Critical, from published rules, with a reason code for every factor. The same facts always give the same level.
Risk classification High
Computed by published rules from the completed baseline assessment.
This level requires a privacy review, a security review and an approved vendor before approval can be requested.
03 Approve
Who approved it, and until when?
- Reviews and approvals. Privacy and security reviews with findings and conditions. Approval with separation of duties, a pinned basis and an end date.
- Vendor governance. Each vendor’s own classification and approval, and every AI system that depends on that vendor.
Support reply assistant
Each review records the reviewer’s decision, with findings, conditions and the next review date.
A privacy review is not a DPIA. A security review is not a certification. The console says so on the page.
04 Control
What is allowed through?
- Admission gateway. A seventeen-stage pipeline: prompt DLP in ten categories, threat detection, data residency, vendor requirements, team budgets, approved models, OIDC and agent signatures.
Policy
Edited in the console. Enforced on every routed request.
05 Prove
Can we prove it?
- Evidence and audit. An append-only, hash-chained log with optional Ed25519 signatures, an integrated status for each subject and a downloadable audit package.
Support reply assistant
One downloadable package for each AI system and each vendor.
The package states what it does not prove.
The product views on this site are illustrative, with invented example data. They show behaviour documented for the current release and are not screenshots of a customer environment.
How control works
Three ways in, one pipeline, and a refusal that never leaves your network
Any licensed, enforcing stage of the pipeline may refuse. After a refusal, every later stage is skipped, the provider is never called and nothing is billed. The refusal names the rule that caused it.
- OpenAI-compatible API
- What an SDK, a coding assistant or an agent framework uses. Streaming is supported.
- Console playground
- For testing and demonstrating a policy before it meets real traffic.
- Proxy capture
- A PAC file and HTTP CONNECT, for traffic that nobody pointed at the gateway on purpose.
- Whole-request scanning
- String content, content blocks, system messages, tool call arguments and tool results.
- Live dispatch
- Anthropic, OpenAI, Azure OpenAI and xAI, plus a self-hosted model endpoint.
Architecture
Where MERIDVAR sits
Compared
The same questions, asked of a cloud-delivered control and of MERIDVAR
| Question | Cloud-delivered AI controls, typically | MERIDVAR |
|---|---|---|
| Where is prompt content inspected? | On the vendor’s infrastructure | On your network |
| Where do governance records live? | On the vendor’s platform | On your volume |
| Is there a new processor in the data path? | Yes | No |
| Does it operate without internet access? | No | Yes, once installed, including licensing and signed updates |
| How is it tenanted? | Multi-tenant | Single-tenant appliance |
This compares typical deployment architectures, based on public vendor documentation. Several vendors also offer customer-operated components. It is not a legal claim.
What we have measured
The product is held to the same standard of evidence
Zero calls on refusal
23 gateway tests were run against a provider that logs every request it receives. Every refusal made zero calls, and every admission made exactly one.
No identifier in the clear
Prompt DLP findings keep masked samples. The raw identifier is not stored.
Offline updates
A signed incremental update was applied to a real appliance with no internet access. Licences were verified on the appliance.
Governance at scale
The status of 5,000 AI systems was computed in about 70 milliseconds.
Deployment model
You run it, and you keep everything it records
MERIDVAR is delivered as software you install and operate. The architecture is the assurance: what never leaves your environment cannot be exposed by ours.
- Self-hosted
- The appliance runs on a host you control, inside your own network.
- Single-tenant
- One appliance serves one organisation. No other tenant’s data sits beside yours.
- Customer-controlled
- Your policies, your keys and your evidence, on your volume. You decide who has access.
- No vendor cloud
- There is no MERIDVAR control plane in our cloud and no new data processor in your AI data path.
- No vendor telemetry
- The appliance sends us nothing. With no integrations configured, it makes no external calls.
- Able to run offline
- Once installed, it runs without internet access, including signed licences and signed updates.
Guides on this topic
See MERIDVAR refuse a request on your own network
A working session with the founder, on a host you control.
What happens next
- You send the formYour request goes straight to the founder.
- We reply by emailWe agree a time and what you want to see.
- We run a working sessionWe install the appliance with you, route a test request through it and read the evidence it leaves.