What an MCP knowledge base server is
MCP is the Model Context Protocol. It is a small, open standard that lets an AI assistant call tools on an outside service. An MCP knowledge base server is one of those services: it holds your notes and exposes a handful of tools, so an assistant can list pages, read a page, search across pages, and write a page back without ever leaving the chat.
The point for a team is simple. Instead of knowledge living inside one person's assistant, it lives in a shared store that every connected assistant reaches. A new chat starts from the same facts the last chat wrote down.
How a team connects one as a wiki
You point each teammate's assistant at the same server URL. Each person signs in once and approves the connection. From then on, their assistant reads and writes the same wiki. You can connect a single wiki, or connect account-wide and let the assistant pick the right wiki for each task. Every connection is scoped to what that person may see, and it can be revoked at any time. The connect an assistant page has the exact steps for Claude, ChatGPT, and Claude Code.
The tools, and the guard that stops silent overwrites
A good MCP knowledge base keeps a small, readable tool surface. agent-wiki ships these:
- list_wikis and list_files — see which wikis and pages exist.
- read_file and search_files — open one page, or search text across a whole wiki.
- write_file — save a page. It uses optimistic concurrency: you pass the version number you got from read_file as expectedVersion. If a teammate's agent saved a newer version first, the write is refused instead of quietly overwriting their edit.
- append_to_inbox — drop a quick note on the wiki's inbox page to file properly later, without stopping to choose a location.
That version guard matters most when several agents write at once. It is the line between a shared store you can trust and a document that loses edits.
Plain Markdown in, plain Markdown out
Every page is plain Markdown. You can export any wiki as .md files whenever you want, so your knowledge is never locked in. Every read and write is logged, every version is kept, and one click reverts a bad edit. The store stays honest: agents only ever work with pages you can open, edit, and export yourself.
agent-wiki as one honest option
agent-wiki is a neutral, cross-assistant knowledge base built for this exact job. It is listed in the Claude Connectors Directory as a community connector, so Claude users can add it without custom setup — and the same wiki also answers ChatGPT, Cursor, and Claude Code. A free tier covers one small wiki; see the home page for plans and limits, and security & trust for how access and privacy work.