Connect an AI agent through MCP

Connect Claude, ChatGPT or any MCP-capable assistant to Nova, as a person or as an AI teammate, and what it can reach once connected.

Nova speaks the Model Context Protocol. Point an assistant at the MCP endpoint and it discovers how to sign in, which tools exist and where the help articles are.

MCP endpoint:          https://app.novabmp.com/mcp
Resource metadata:     https://app.novabmp.com/.well-known/oauth-protected-resource/mcp

Two ways an assistant can act

• As a person. In Claude or ChatGPT, add Nova as a custom connector with the endpoint above. The assistant registers itself, sends the person to Nova, the person presses Allow, and the assistant acts as that person with the scopes they approved. The person can disconnect it any time under Profile › Connected apps. On the same screen, under Connect as, a person who looks after an AI teammate (or an admin) can hand the assistant that teammate instead of themselves; it then acts under the teammate's name and roles.
• As an AI teammate. A person in the organization adds it under Administration › Team & Access › AI Teammates, gives it roles like any team member, and issues a secret. Your AI platform signs in with client credentials (Run software as an AI teammate). Everything it does is recorded under its own name, with its accountable person on record.

What the assistant gets

• Documented tools with the same names as the operation ids in API reference: list_accounts, create_quote, send_invoice, search_help, get_help_article and so on. A tool is listed only when the connection holds every permission it checks.
• Everything else through find_operations and call_nova_operation, which search and call any of Nova's operations the connection's grant covers.
• Built-in tools: whoami and list_my_organizations.
• Resources: every help article as nova://help/<key>, so the assistant can read how a screen works before it acts.
• Prompts: short recipes such as creating a quote for a customer.

The full list with examples: The tools an assistant gets over MCP.

Approvals

The few actions Nova guards with an identity check (payout accounts, role changes and member removal, bulk exports) always wait for a person. When an AI teammate signs in with its own credentials, the actions its organization listed under Ask a person first wait too. A connector added in Claude or ChatGPT has no such list of its own, including when it connects as a teammate. When the assistant attempts one of these actions, the call returns 403 approval_required with an approval_request_uid. A person approves it in Nova's Approvals inbox; the assistant then repeats exactly the same call and it runs once. Approvals last 72 hours. Details: Handle an action that waits for a person's approval.

Limits that always apply

• A grant narrows, never widens: the assistant never does more than the person it acts as, or an AI teammate's roles, allow.
• Paused or retired AI teammates lose their tokens immediately.
• Per-minute limits per token and per organization: Rate limits, daily caps and other limits on a credential. Back off on 429. An AI teammate's working hours apply however it connects; its daily cap applies when it signs in with its own credentials.
• A customer-facing AI teammate's texts and emails end with an AI disclosure line that Nova adds.
• Every call is in the audit log (Administration › System › Audit Log, the API & Agents view), and an AI teammate's own activity is on its card.

Testing without an assistant

Use the same tools by hand:

POST /mcp
Authorization: Bearer <token>
Content-Type: application/json

{"jsonrpc":"2.0","id":1,"method":"tools/list"}

Then call one:

{"jsonrpc":"2.0","id":2,"method":"tools/call",
 "params":{"name":"list_accounts","arguments":{"q":"smith","limit":5}}}