Add the protection to the product your customers use.
Building an application, a portal, a shop or an AI service? Connect Corpilus checks where your product takes in content, works with data and talks to AI. We design the scope and the way it is integrated together.
Integrations are just as useful for protecting your own company application — they are not only for partners. An SDK or an API is not another protection package; you are connecting the same Corpilus. How demanding the implementation is can be reflected separately in the quote.
Pick the specific checks for your product.
During the integration we settle what you hand over for checking, what result your application receives and how it reacts to it. That way you can add protection to one specific user flow and verify it on your own scenarios.
A scope you can read
We agree which flows and which kinds of content go for checking. The scope is named in the integration, not inferred from the name of a package.
An agreed point of intervention
For every flow we settle whether a finding is only recorded or stops the request, and who enforces the decision — your own handler or the connected interface.
Verification before production
We go through the scenarios on your own requests: the ordinary course, a finding, a wrong finding and a check that fails.
What you want to achieve, and what connects it.
Every row is a way to connect that exists today. The third column is what we cannot design an integration without.
| I want to achieve | Way to connect | What has to be agreed |
|---|---|---|
| Check web forms and selected requests | Widget and server-side integration (SDK) | Which routes verify the result and what they do on an error. Ready-made middleware exists for Python, Node.js and PHP including Laravel, plus an add-on for WordPress and WooCommerce. |
| Evaluate content from your own application | REST interface | The supported kind of input, the check available for it and how your handler reads the result. |
| Check the traffic to a supported model | Proxy in front of the model | Addresses, keys, streaming and the reaction to a finding. We can parse and refuse the OpenAI and Anthropic request shapes. |
| Call protective tools from an AI client | MCP checking tools | The difference between a check you ask for and a filter over the whole flow. |
| Check tool calls in a connected gateway | MCP gateway | Which calls pass through the gateway and where the reaction is enforced. |
| Protect employees' work in a supported chat | Browser extension | Compatibility, sign-in and the mode of checking. ChatGPT, Claude and Gemini are supported. |
| Connect the protection to infrastructure of your own | Individual integration | Technical feasibility, the scope of the work and operational responsibility. |
Running on your own server decides where the protection runs. It does not replace connecting the specific flows of an application.
Deployment and environmentsFrom one flow to a verified integration.
- 01You describe the product and the need.What the user does, and which data or requests have to be checked.
- 02We propose how to connect it.The interface, the data flow, the reactions and the operating conditions.
- 03We verify the scenarios.The ordinary course, a finding, a wrong finding and a check that fails.
- 04We agree the operation.Scope, capacity, price, support and responsibilities.
A demo, a sandbox and a pilot are not offered publicly; the terms, price included, are agreed beforehand. Calling something a pilot does not imply free operation or full production scope.
The same fields for every interface.
The same for every way to connect: what you hand over, what comes back, what authenticates the call, where the decision applies and what happens on an error. A code sample appears only where it could be verified against the SDK as you download it from the dashboard and the integration snippets.
Widget and server-side integration (SDK)
- Input
- The script on the page hands over signals about the work with the form; your handler hands over the submission token and the named fields.
- Output
- A signed token for the browser, and a decision for your handler.
- Authentication
- The site key in the script, the interface key for the server-side call.
- Point of intervention
- In the form handler on your server. The script on the page enforces nothing on its own.
- Error states
- A token that could not be verified is recorded as missing and never refuses the submission.
- Input
- The script on the page hands over signals about the work with the form; your handler hands over the submission token and the named fields.
- Output
- A signed token for the browser, and a decision for your handler.
- Authentication
- The site key in the script, the interface key for the server-side call.
- Point of intervention
- In the form handler on your server. The script on the page enforces nothing on its own.
- Error states
- A token that could not be verified is recorded as missing and never refuses the submission.
REST interface
- Input
- The content you send for checking, or a token to verify.
- Output
- A finding, or the result of the token check, in the response to the call.
- Authentication
- The site key and the interface key in the request headers.
- Point of intervention
- In your application: the interface answers, your handler decides.
- Error states
- A failed call is an outage of your own call — nothing is refused on our side.
- Input
- The content you send for checking, or a token to verify.
- Output
- A finding, or the result of the token check, in the response to the call.
- Authentication
- The site key and the interface key in the request headers.
- Point of intervention
- In your application: the interface answers, your handler decides.
- Error states
- A failed call is an outage of your own call — nothing is refused on our side.
Proxy in front of the model
- Input
- The prompt you send to the provider, and its answer.
- Output
- The request carries on, or you get an error with a reference instead of an answer.
- Authentication
- The site key in a header; your own key with the provider passes through unchanged.
- Point of intervention
- Before the request goes to the provider, and as the answer comes back. A streamed and a non-streamed flow behave differently.
- Error states
- A body that cannot be parsed carries on and the failure is recorded.
- Input
- The prompt you send to the provider, and its answer.
- Output
- The request carries on, or you get an error with a reference instead of an answer.
- Authentication
- The site key in a header; your own key with the provider passes through unchanged.
- Point of intervention
- Before the request goes to the provider, and as the answer comes back. A streamed and a non-streamed flow behave differently.
- Error states
- A body that cannot be parsed carries on and the failure is recorded.
MCP checking tools
MCP gateway
- Input
- The tool name and the arguments of every call that passes through the gateway.
- Output
- The result of the call, or a refusal with a reason.
- Authentication
- A key of its own for the MCP interface, in the Authorization header.
- Point of intervention
- Before the tool runs.
- Error states
- When the classification fails, a call in a risky class is held and the rest carries on.
- Input
- The tool name and the arguments of every call that passes through the gateway.
- Output
- The result of the call, or a refusal with a reason.
- Authentication
- A key of its own for the MCP interface, in the Authorization header.
- Point of intervention
- Before the tool runs.
- Error states
- When the classification fails, a call in a risky class is held and the rest carries on.
Browser extension
- Input
- The text before it is sent in a supported tool, and the attachments added to it.
- Output
- The text according to the chosen mode: a warning, substituted values, a rewrite or a stop.
- Authentication
- The employee's sign-in to the workspace.
- Point of intervention
- In the browser, before the text leaves.
- Error states
- A file that could not be read is reported as unchecked with the reason, never as clean.
- Input
- The text before it is sent in a supported tool, and the attachments added to it.
- Output
- The text according to the chosen mode: a warning, substituted values, a rewrite or a stop.
- Authentication
- The employee's sign-in to the workspace.
- Point of intervention
- In the browser, before the text leaves.
- Error states
- A file that could not be read is reported as unchecked with the reason, never as clean.
What exactly is checked in each channel, what does the checking and what the operating limits are is on the page for IT — it is not repeated here.
For IT: the scope of the checks and the limits