MCP

probr runs a Model Context Protocol server, so an AI assistant can read your workspace directly — your competitors, your recent signals, your latest brief — and answer questions about them without you opening anything.

This is on every plan, and on the trial.

The address#

https://mcp.probr.ai/mcp

That is the whole configuration. There is no API key to generate and none to paste. The first time a client connects it opens your browser, you sign in to probr and approve the request, and the client holds the token from then on.

Claude Code#

claude mcp add --transport http probr https://mcp.probr.ai/mcp

The first tool call opens a browser for sign-in and approval. After that it is connected.

Claude Desktop#

Settings → Connectors → Add custom connector, with the same URL.

Cursor#

Cursor reads ~/.cursor/mcp.json for servers available everywhere, or .cursor/mcp.json inside a project for one repository:

{
  "mcpServers": {
    "probr": {
      "url": "https://mcp.probr.ai/mcp"
    }
  }
}

Restart Cursor, then open Settings → MCP. probr will be listed as needing authentication; approving it opens the browser sign-in.

Any other client#

If it speaks MCP over Streamable HTTP and supports OAuth, it will work — the server implements the discovery documents clients look for, so there is nothing probr-specific to configure. Point it at https://mcp.probr.ai/mcp and let it find the rest.

probr is also listed in the official MCP Registry as ai.probr/probr, so a client that browses the registry can find it without being given the URL.

If you are writing the client yourself, OAuth has the endpoints.

What you can ask#

Twelve tools, described properly in MCP tools. In practice they turn into questions like:

• What have my competitors done in the last two weeks?
• Has Globex changed anything about pricing recently?
• Summarise the latest brief and tell me what to worry about.

The assistant is reading the same signals the feed shows. It is not doing new research, and it cannot see anything your workspace cannot.

Attachments and commands#

Two things clients offer that are not tools, because you pick them rather than the model:

Resources are documents you attach to a conversation — this week's brief, your tracked list, or one competitor's last 90 days. Attaching the brief and asking about it is more direct than hoping the assistant decides to look it up.

Prompts appear as commands in the client's menu: a weekly review, a triage of critical signals only, a sales battlecard for one competitor, and a catch-up over a window you choose. Each arrives with its data already fetched.

Some clients also render an interactive board in the conversation — today's signal tally by severity, with the signals behind it. Clients without that support get the same answer as text.

What it can change#

Reading is the default and writing is separate. A connection is read-only unless you approve the probr.write scope on the consent screen, and a client that does not hold it is not shown the write tools at all — they are not listed, so there is nothing for an assistant to try.

With that scope, four tools exist: add a competitor, pause one, resume one, and stop watching one. Adding and removing both ask you first, in the conversation, naming the competitor — because the host's own "allow this tool?" prompt names the tool and not which competitor is about to go, and a host can be told to allow a tool once and stop asking.

Nothing touches delivery settings or billing at any scope.

It is scoped to one workspace — the one you were signed in to when you approved. A token issued for your workspace cannot read anyone else's, and the server re-checks that on every call, not just at connection.

When it will not connect#

"This workspace has no active plan." The trial ended, or a subscription did. MCP is included in every plan, so choosing one in Settings is the whole fix — nothing about the client configuration needs to change.

The browser opens and nothing happens after you approve. The client is waiting on a redirect back to itself. This usually means the client was not running on the port it registered, which is fixed by starting the connection again from the client rather than reloading the browser page.

Tools stop working after a while. Access tokens last an hour and clients refresh them silently. If a client does not, disconnect and reconnect it once.

406 Not Acceptable, from a client you wrote yourself. Every request has to offer both content types:

Accept: application/json, text/event-stream

Both, literally — */* is refused, and so is the default many HTTP libraries send (application/json, text/plain, */*), because neither names the event stream. The protocol requires it because a response may arrive as either.

405 on GET /mcp. There is no server-opened stream to attach to: probr holds nothing between requests, so every answer comes back on the request that asked. A client that treats the 405 as fatal is being stricter than the protocol, which makes that stream optional.