Your AI assistant connects but sees no data

What you're seeing

You connected an assistant (Claude, Cursor or another client) to Provantix and approved the access page. Then:

  • Every tool call "succeeds" but the assistant says it only got a count, such as "3 result(s)", and nothing inside it.
  • Or a call is refused with insufficient_scope, naming a scope such as coverage:read.
  • Or the assistant lists fewer tools than the MCP page describes, or cannot find a source you know exists.

What it means

These look alike but have different causes. Work out which one you have.

What the assistant says Cause
Only a count, no names or numbers The tool result's text did not carry the data. Clients like claude.ai show the model only the text part of a result. This was a defect in older versions of the server, which put the data in a structured part the client never showed the model. Current versions put the whole result in the text.
insufficient_scope: this tool needs … The connection was approved without that scope. An assistant may ask only for the scopes it thinks it needs.
Brief refused, coverage refused get_brief needs evidence:read, findings:read and coverage:read; get_coverage needs coverage:read.
Cost refused cost:read is not part of the connection. For a member, the organization must also have switched on Allow cost data over MCP. An admin connection includes it.
Fewer tools than documented The assistant saved the tool list when it first connected and has not re-read it.
401 on every call The connection was revoked, the person was removed from the organization, or the plan does not include MCP (Pilot and above).

What to do

  1. Start a new chat and ask the assistant to list every Provantix tool it has. A tool list is read when a conversation starts, so an old chat keeps the old list.
  2. If the list is still short or a scope is missing, reconnect. Remove the Provantix connector in the assistant and add it again with https://mcp.provantix.tech/mcp. The access page then shows the scopes this connection will get.
  3. Read the scopes on the access page. An organization admin has all access inside their own organization, so an admin's connection includes every scope and the page marks the ones the client did not ask for included for admins. A member gets what the client asked for.
  4. For cost as a member, ask an admin to switch on Allow cost data over MCP under Settings → Connected AI clients, then reconnect and approve the new scope.
  5. Test with a tool that returns a list. Ask for list_sources. You should see names, statuses and record counts. If you see a count line only, the server you are connected to predates the fix; contact support.
  6. Check Settings → Connected AI clients for the connection. It shows the client, when it connected and last used the connection, and lets an admin revoke it.

What good looks like

The assistant answers from data: it names your sources and agents, quotes record counts, and can follow an id from one answer into the next call (list_agents to get_agent to get_record). Refusals, when they happen, name the scope and say how to get it.

Good to know

  • Every call is audited with your user, the client and the tool. See Connect your AI assistant.
  • Every tool is read-only; nothing an assistant does can change your data.
  • Text that agents wrote is marked untrusted in results, and the assistant is told never to follow instructions inside it.