Grodmir — AI gatewayYour data.
Your boundary.

A self-hosted AI gateway checks requests from connected AI tools before they reach a model provider. Inspect sensitive data locally, apply signed policies and control what leaves your infrastructure.

For routed channels. Observe mode forwards content unchanged.

The boundary, made visibleIllustration · not live traffic

01 Architecture

The boundary lives
on your side.

Inspection happens inside your infrastructure. The EU control plane receives metadata, not the content of your prompts.

Your infrastructure

People & tools

Prompts, code and tool results

Grodmir

Inspect. Apply policy. Audit.

Model provider

Content allowed by your policy

A policy can get stricter.
It can’t quietly get weaker.

Organisation, end-client and project policies form a hierarchy. A narrower scope can tighten the rules. A policy exception must be explicit; secrets cannot be exempted.

The gateway sees raw content. The model sees what the effective policy permits. The control plane sees metadata only.

OrganisationEnd clientProject

Read how a self-hosted AI gateway controls routed traffic.

Next: Channels

02 Channels

Three ways in.
One policy boundary.

Employees, applications and agents use different tools. Protection applies when their traffic goes through a connected channel.

A /

Applications & coding assistants

API gateway

OpenAIAnthropicClaude CodeCodexCursor

Checks model requests routed through the gateway and applies the effective policy.

Outside its view: direct requests, personal provider keys and tools that bypass the gateway.

Connect itDeploy the gateway, configure its policy and keys, then change the tool’s base URL and credentials.

AI tool integration guides

B /

Pasted text in browser chats

Browser extension

Paste & dropYour gateway

Sends pasted or dropped text to your gateway for inspection before the page receives it.

Outside its view: typed text, other upload paths and browsers without the extension.

Connect itDeploy the extension, point it at your reachable gateway and provide an inspection-scoped key.

Browser extension rollout and coverage

C /

Agents & external tools

MCP proxy

Tool argumentsTool results

Checks arguments on the way to an MCP server and results on the way back into the model context.

Outside its view: MCP servers connected directly, without the proxy.

Connect itInstall the proxy, configure a gateway inspection key and wrap each intended MCP server connection.

MCP proxy setup and protection limits

Next: Get started

03 Get started

First understand.
Then enforce.

Start with observation. Measure the effect on your own data before relying on the policy.

  1. 01

    Observe the traffic

    Run connected channels in observe mode. Content is forwarded unchanged; policy blocks and transformations are not applied.

    EvidenceMetadata audit entries for inspected requests.

  2. 02

    Read the report

    Review detected categories, decisions and uncovered channels. Check misses and false alarms on your own examples.

    EvidenceA risk report based on metadata counters.

  3. 03

    Set the policy

    Define organisation rules, then refine them for each end client and project. Review and publish a signed version.

    EvidenceVersioned, signed policy and publication history.

  4. 04

    Enable enforcement

    Publish the enforcing configuration and verify it on your connected channels. Monitor actual request decisions in the local audit.

    EvidenceLocal audit decisions; signed end-client attestations on supported plans. Attestations alone do not prove device adoption.

Bring your organisation.

Create your workspace to manage a subscription and access setup resources. A gateway deployment is still required.

Use the AI gateway selection checklist to plan your pilot.

Next: Policy

04 Policy

Same message.
Different exposure.

An identifier becomes a token. A secret is removed. A confidentiality rule can block the whole request. The stricter decision wins.

Illustrative policy · synthetic data only

01 / Before inspection

Contact jan.kowalski@example.com.
API key: SYNTHETIC_API_KEY_EXAMPLE.
Use this for the project summary.

02 / After the example policy

Contact [EMAIL_1].
API key: [SECRET_REMOVED:API_KEY].
Use this for the project summary.

Transformed · permitted by this example policy

The model receives the token and removal marker, not the original values. This illustration does not run the production detector.

See how Prompt DLP handles sensitive data and secrets.

Next: Your data

05 Your data

What sees your data.
And where it goes.

The browser extension, the MCP proxy and the gateway read your content to protect it. All three run in your infrastructure and send content only to your own gateway. We, the vendor, never receive it.

The vendor-hosted control plane is not currently deployed; this diagram describes the product architecture.

ComponentRuns onSees your content?Sends content toResponsible for the content
Browser extensionYour employees’ browsersYes: pasted and dropped textOnly your gateway, set by your browser policyYou (controller)
MCP proxyYour workstations or serversYes: tool arguments and resultsOnly your gatewayYou (controller)
GatewayYour infrastructureYes: everything it inspectsThe AI models your policy allows, after tokens and removalsYou (controller)
AI model providerThe providerOnly what your policy lets through—You, with the provider as your processor
EU control planeSelf-hosted; vendor hosting not yet selectedNo: counters, categories, policy versions, administrator e-mails—Metadata processor when operating the control plane (DPA)
This websiteRailway (EU West)No: account and billing data onlyRailway hosting; Clerk sign-in when configured; Stripe paymentsWe are the controller of account data
01 / Never

What we never receive

Prompts, answers, files, tool results, detected values, pseudonym maps, your audit log and your model-provider keys.

02 / Locked

Nobody can redirect it

The production extension takes its gateway address only from your browser policy. Employees cannot point it at another server.

03 / Verify

Check it yourself

Watch the extension’s network traffic, block our domains at your firewall, or send a synthetic canary and search your control-plane export for it.

04 / Support

Support uses synthetic data

We reproduce problems with synthetic examples and never ask for real prompts, files or screenshots with content.

05 / Your duties

Inform your employees

You decide what is inspected. Tell your team before rollout; in Germany, involve the works council. We provide notice templates in PL, DE and EN.

06 / Roles

When roles would change

If we ever host a gateway for you, we would see content and need a separate data processing agreement. That is not part of the standard offer.

Review the AI gateway data boundary and security limits.

Next: Limits

06 Limits

A clear boundary.
Including its limits.

Grodmir reduces exposure in supported channels. Know what is outside that scope before depending on it.

01 / Content

How are files and images inspected?

Files are inspected only when attachment inspection is enabled by policy. Image OCR also requires an OCR-enabled build. Unreadable content is handled by policy; the default is to block. Files are never rewritten: a required removal or pseudonymisation blocks the request.

02 / Language

Text in PL, DE and EN

Detection supports Polish, German and English. Voice and other languages are outside the scope.

03 / Coverage

Connected channels only

A path that bypasses the gateway, extension or MCP proxy is not protected. The browser extension covers paste/drop, not typed text.

04 / Detection

Validate on your own data

The model for names, organisations and addresses was trained on synthetic data. Quality on real documents is not validated; misses and false alarms occur.

05 / Verification

No external penetration test yet

Do not treat the product as independently security-audited. Browser coverage depends on each provider’s current page layout and must be checked after changes.

06 / Intent

Protection against mistakes

Deliberate obfuscation is not reliably detected. A local administrator can disable the gateway. Network controls remain part of your deployment.

Read the full protection limits before deployment.

Next: Who it’s for

07 Who it’s for

When your client asks,
show the control.

For software houses of 20–500 people serving clients in Germany, Austria and Switzerland. When the conversation turns to NIS2, GDPR or ISO 27001, Grodmir provides policy and audit evidence about controlled AI channels. That evidence supports the discussion; it is not a certification or a guarantee of legal compliance.

Starting in Poland. Built with the DACH direction in mind. Detection in Polish, German and English.

Next: Get started
Grodmir.