AI agents working within the perimeter

The MCP server exposes 34 tools with the minimum required role and a read-only or destructive annotation: flows, cases, tasks, documents, events, dashboard and analysis, Process Map and Solutions. It uses Streamable HTTP transport, enforces schemas before any change and supports safe retries through expected state, idempotency keys and data revision.

For ChatGPT and claude.ai the connection goes through OAuth. The Admin creates a profile with maximum role, tools and flows; the client registers itself and the user approves a consent screen showing exactly what is granted. PKCE is mandatory, effective access is the minimum between profile and current permissions, and revocation is immediate. MCP keys have an expiry preset at 90 days and never travel in the query string.

Behind every action there is a person. History shows the AI client and the key holder, and the AI cost of a case started by an agent is attributed to the key. The built-in Flow Agent, always at hand with a keyboard shortcut, works within the user's permissions and applies Data Shield. Map and solution imports from agents are preview by default, and applying requires a request identifier and explicit consent.

A typical case: an assistant connected via OAuth reads late cases, proposes starting a new one for a client and fills in the initial data. The case is born waiting; a person confirms or rejects it with a reason. History shows the assistant, the person holding the connection and the AI cost attributed to the key, so the owner knows who produced what.

For IT, the value is predictability. Schemas are enforced before any change, retries are safe thanks to idempotency, and errors return their cause. For the compliance officer, the value is readability: keys with expiry, profiles with a maximum role, consent shown to the user and immediate revocation. The agent becomes a collaborator with a job description, not a forgotten privileged access.