| Where people use it |
Inside RR, in the browser, beside the rules they are looking at |
In their own AI client: Claude on the web, desktop or phone, an AI coding tool, or your own agents |
| Who runs the model |
Your server, through the client you register (Anthropic, OpenAI, Azure OpenAI, a model on your own hardware, any IChatClient) |
The AI client's vendor, with the model the person or their organization chose there |
| Who pays for the model |
You, on your vendor account, per token. You can set prices, see usage and enforce quotas (OnUsage, CanStep) |
Whoever pays for the AI client (usually each person's or your organization's subscription). Your server pays nothing per token |
| Where the data goes |
To the vendor you chose, under your contract, or nowhere if the model runs on your hardware |
To the AI client's vendor, under the person's or organization's agreement with it |
| Sign-in |
None extra: the person is already signed in to your application, and RR's endpoint uses that |
OAuth through your identity provider (Microsoft Entra ID, Okta, Auth0 ...). Each AI client is registered with it, and the server must be reachable from the AI vendor's cloud when the client runs there |
| What you build |
A few lines: register the models, set Assistant on the RR configuration |
An ASP.NET Core (.NET 8) endpoint, token validation, the protected resource metadata, a mapping from the token to a user and permissions |
| Runs on |
Wherever RR runs: .NET Framework or .NET |
.NET 8 or later, ASP.NET Core only |
| Needs RR's user interface |
Yes: it lives in RR's screens |
No: works with the Rule Repository library, Audit, or both, with or without RR's screens |
| Tools |
The same tools, loaded in groups as the conversation needs them; the AI Rule Editor has a fixed set that saves nothing |
Every tool the caller's permissions allow, listed per caller |
| Context |
Knows where the person is: the AI Rule Editor starts with the rule, draft or version on screen |
Starts from nothing; the person names the rule or source |
| Confirmations (a test that calls production code) |
A card in the conversation; the person answers there |
The client asks the person if it supports it; otherwise the agent must ask and call a separate confirming tool |
| Files |
Attachments in the AI Assistant (PDF, images, text), kept where you decide |
Whatever the AI client supports; the client reads them, not your server |
| History |
Conversations stored in RR's database for 90 days by default; administrators with permission 112 can read every user's |
Kept by the AI client, not by your application |
| Activity |
Every save names the person and the agent (by default "AI Assistant") |
Every save names the person and the AI client the token was issued to |
| Automation with no person |
No |
Yes: a service or an agent of your own can sign in with client credentials or a key and run, for example, nightly test runs |
| Skills |
Built in: RR loads the skill for each group of tools |
Served by the server; clients that load skills over MCP get them, and the server's instructions carry the essentials for the rest |