The MCP server for AI agents

Read and draft, never enforce. The server at https://app.pubtrust.cc/mcp lets an agent query metrics and trends, drill into violations, inspect sites, policies and packs, and generate reports. It can also author policy: create and edit policies, add keywords, domains and URL patterns, propose terms back into a pack and create fallback creatives. It cannot switch a policy from Monitor to Block, and it cannot touch billing.

Register the server over HTTP, passing an API key from your PubTrust account as a bearer token:

claude mcp add --transport http pubtrust https://app.pubtrust.cc/mcp --header "Authorization: Bearer <key>"

Replace <key> with the key. The server speaks Streamable HTTP; clients that support OAuth can connect to the same URL without a key, using OAuth 2.1 with PKCE, in which case the person authorising the agent signs in to PubTrust and approves the connection.

Two: discover and execute. discover returns the operations available to the connected credentials, with their input schemas; execute runs one operation by name with arguments that match its schema. Keeping the tool list to two keeps an agent's context free of dozens of tool definitions, and lets the agent find the operation it needs when it needs it.

Enforcement and billing. When an agent recommends moving a policy from Monitor to Block, it creates an enforcement request, which appears as a pending approval in the dashboard; nothing changes until a person approves it there. Billing — plans, payment methods, invoices — is outside the agent's scope entirely.

Every MCP action is written to your organisation's append-only audit log with the agent's identity recorded separately from the person who authorised it, so a change an agent made and a change a person made remain distinguishable afterwards.