To build a company MCP your whole team can use inside Claude: import your website and upload your internal documents into one knowledge base, publish it as a private MCP server, and have each teammate connect it in Claude once. Sign-in runs on OAuth 2.1, the server is isolated to your workspace, and from that point everyone's Claude answers company questions from the same cited source — not from memory, guesswork, or whoever sits closest to the person who knows.
The pattern this replaces is worth naming. Your team already uses Claude, but Claude does not know your company: not your SOPs, not your pricing rules, not the way you actually run projects. So people paste documents into individual chats, get answers of varying vintage, and the real knowledge keeps living in two colleagues' heads.
If the answer lives in one person's head, it is not an asset. It is a queue.
What belongs in the company brain
- SOPs and runbooks — how things are actually done here
- Policies: refunds, cancellations, security, leave
- Product facts — specs, limits, versions, pricing
- Onboarding material, so new hires can ask instead of interrupting
- The internal FAQ nobody maintains but everybody needs
How teams feed Claude company knowledge today
The first two are not wrong so much as unaccountable. Nobody knows which version anyone is working from, and when an answer is bad there is nothing to trace.
The rollout
Build it once
Import the website — crawled into a cited knowledge base in about two minutes — and upload the internal documents the site would never contain. No code in any of this.
Keep it private
Private is the setting that makes it a company brain: your team, authenticated, and nobody else. Public and even paid-access servers exist on the same platform, but that is a different project.
Publish and share the endpoint
Publishing produces the MCP server. Send the connection details to the team the way you would share any internal tool.
Everyone connects in Claude
Each person adds the server and signs in through OAuth once. After that it is simply there — Claude can search the company brain mid-conversation and cite what it found.
The security paragraph, since someone will ask
Published servers implement OAuth 2.1 as the current MCP spec defines it, including issuer validation and CIMD client registration, and every server is isolated per workspace. In practice: your people sign in, the auth is the modern spec rather than a homemade approximation, and your workspace's knowledge is reachable only through your workspace's server. This is the part that quietly kills most build-it-ourselves internal tools, and here it is already done.
Not only Claude
MCP is a protocol, not a Claude feature. The same private server works in any MCP client, so a teammate who lives in a different tool connects to the same brain and gets the same cited answers. You standardize the knowledge without standardizing everyone's tooling. The same build also produces a chat agent, if part of the team prefers a plain chat window over an MCP connection.
Week one
Seed it deliberately: the ten questions your team actually asks, answered by the documents you upload. Then let people try to break it. Wrong answers are productive here — every answer carries a citation, so a bad one points at the precise document that needs fixing, and fixing the document fixes it for everyone at once. Website-sourced facts stay current on their own, re-syncing as often as every fifteen minutes.
There is a seven-day trial, with plans priced in EUR. A realistic first milestone: by the end of week one, the newest person on the team gets a better answer about your own processes from Claude than from the colleague who has been there five years. Cited, too.