What is an MCP server?
The Model Context Protocol (MCP) is an open standard for connecting AI apps to outside systems. An MCP server sits in front of one system, such as a CRM or a sales intelligence tool, and offers the AI app a short list of named tools it can call when a question needs that system's data.
For a sales team, an MCP server means asking Claude or ChatGPT a question in plain words and getting the answer from your own data, without exporting a spreadsheet first.
What a sales team can ask an AI app once its data is connected
Once a sales system is connected over MCP, a rep can ask the AI app what they would ask a colleague:
- Which accounts should I work this week, and why?
- What changed at this company this month?
- Who is on the buying committee there, and who on our team can introduce me?
- Give me a brief on this account before my call.
- Which competitor is in place there, and how do we win against it?
- Which companies had a new signal since Monday?
The AI app writes the answer. The server decides what the app can see and do, so that is where to look before connecting sales data.
What to check in any vendor's MCP server
Six checks apply to any MCP server that touches sales data.
- Read or write. List every tool the server offers. A read-only server answers questions. A server with write tools can send an email, change a CRM field or enrol a contact, in your name. Check that read-only tools are marked read only, so the AI app treats none of them as a change.
- Sign-in. Each person should sign in with their own account, through OAuth, so access follows what that person can already see. Where a shared key is used, ask how long it lives, who can revoke it, and whether revoking stops it at once.
- Which data leaves. Ask which fields each tool returns, especially names, titles, emails and phone numbers, and whether every field is listed in the vendor's privacy policy. What the AI app's own vendor keeps follows that vendor's terms and your settings.
- Per-board consent. At sign-in, you should be able to choose which boards, workspaces or lists the app may read.
- What the vendor logs. Ask whether the vendor keeps your questions and answers, or only a record of each call.
- Who switches it on. Access should be on or off per workspace, with an admin who can see every key and revoke any of them.
noombat's MCP server
noombat's MCP server, at https://mcp.noombat.ai, lets an AI app read the noombat boards you allow: the companies to work this week and why, the people there, the dated signals and the competitors. It reads and never writes.
- Apps. ChatGPT, Claude (on claude.ai and the desktop and phone apps), Codex, Claude Code, Cursor and VS Code.
- Transport. Streamable HTTP.
- Sign-in. OAuth 2.1 with PKCE. You sign in with your noombat account, tick the boards the app may read and choose Allow. The connection lasts until you disconnect it.
- Keys for code tools. An admin makes an organisation key under Settings, API & AI apps, sent as
Authorization: Bearer nbo_…. A key lives 90 days, a year, or never, and reads one board. - Limits. 600 calls a minute and 10,000 a day per connection or key.
- Logs. One line per call, kept for 13 months. noombat never keeps the question or the answer.
- Control. Access is switched on per workspace. Disconnect an app under Settings, API & AI apps, and it stops at once.
The nine tools on noombat's MCP server
| Tool | What it reads |
|---|---|
| search | Finds a company, person or competitor by a word of its name or a web domain |
| fetch | One company, person, signal or competitor in full, with a link to it on the board |
| list_companies | Companies in the order to work them, best first, filtered by score, flag, segment or what changed since a date |
| get_company | One company's file: profile, score, the dated signals behind it, the people there and the competitors in the way |
| list_people | People at the board's companies, with business title, seat in the buying decision and public profile link |
| get_person | One person: business title, seat, public profile link and the do-not-contact flag |
| list_signals | Dated events about the companies, newest first, each with a link to the page that published it |
| list_competitors | The competitors on the board, each with how you win against it |
| get_competitor | One competitor: its name, its domain and how you win |
Every tool is marked read only. None sends a message, writes to a board or touches a CRM.
noombat's MCP server puts the ranked companies, people, signals and competitors from the boards you allow inside the AI app your team already uses, read only: see the MCP server page or Book a demo.
Questions
Is it safe to connect an AI app to sales data over MCP?
It depends on the server. Check whether any tool can write or send, how people sign in, which fields leave, and whether you choose what the app may read. noombat's server is read only and reads only the boards you tick at sign-in.
Which AI apps can use the noombat MCP server?
Any app that speaks MCP over HTTP: ChatGPT, Claude, Claude Code, Codex, Cursor and VS Code. In ChatGPT and Claude you add it as a custom connector or MCP server and sign in. Whether an account can add one depends on its own settings and workspace.
Can the noombat MCP server send emails or update our CRM?
No. Every noombat tool reads, and each one is marked read only. The connection cannot send a message, write to your CRM or change your board.
What does noombat keep when an AI app reads a board?
One line per call: who, which app, which tool, when, how many rows and whether it worked. noombat keeps it for 13 months and never keeps the question or the answer. What the AI app keeps follows OpenAI's or Anthropic's terms and your own settings.
How long does it take to connect an AI app to noombat?
About five minutes, for a member of a workspace where AI apps are switched on. In Claude Code, run claude mcp add --transport http noombat https://mcp.noombat.ai, then /mcp to sign in. Code tools that take a header can use an organisation key instead.