Guide · · 2 min read
MCP servers for team knowledge: why tool access isn't enough
Claude Code, Cursor and Codex reach your tools but not your project. How an MCP server connects team knowledge, keeping sources and permissions.
Niclas GregorCo-founder of Scibly
The key points at a glance
- MCP is an open standard through which Claude, ChatGPT, Cursor, Copilot and other assistants call a server's tools.
- One MCP server per tool gives the agent access, not connections: it does not know what means the same, what's current and what was decided on a call.
- A server for team knowledge reads your sources ahead of time, links them and names the source of every statement.
- Everyone signs in themselves and gets only what they may see in the sources.
What is MCP?
The Model Context Protocol is an open standard Anthropic introduced at the end of 2024 (specification). An AI assistant uses it to call a server's tools, such as "search", "read this document" or "create a ticket". An assistant that calls such tools itself is called an agent.
https://app.scibly.com/api/mcp
Why is one server per tool not enough?
By now almost every tool has its own MCP server: with GitHub's the agent sees issues and pull requests, with Notion's pages, with Atlassian's tickets. For each question it then searches every tool live, with search words it makes up each time. Here is the same question with both kinds of server:
When does Project Atlas start, and who decided that?
14 issuesATLAS · 23 ticketsStart Sept 1Planning call, not read“Atlas starts on September 1.”Out of date. Who decided stays open.
holds nowsaid in the calldecidedProject Atlasatlas-webATLAS · 23 ticketsStart Sept 1Start Oct 15Mia Roth“Atlas starts on October 15; Mia Roth moved it.”Right, with its reason and source.
The agent is missing three things:
- Connections. No server knows that "Project Atlas" in the ticket, "the new customer portal" in the meeting and "atlas-web" in the repository are the same thing.
- What's current. The search finds the old statement and the new one alike, and the agent has to guess which is right.
- What was said. Many decisions are made on a call and never written down. No tool server has them.
What does a server for team knowledge have to do?
- Read and connect in advance. It reads your sources before anyone asks and links the people, projects and decisions in them, even when they go by different names.
- A source for every statement. The agent gets the document, email or meeting it came from, so a person can check it.
- Only what the person may see. Everyone signs in themselves, and the server answers only with what they could open in the sources. One shared login for everyone would show everyone everything.
- Writing back. The agent can keep what it learned, such as a team rule, with the person as its source. Next time, every agent finds it.
What this looks like in Scibly
Scibly, our product, is such a server. It reads your meetings, the tools each person connects themselves (Google, Notion, Jira, Confluence, Slack, GitHub), and what people and their agents save. How it handles that data is under Security. Through it, an agent gets these tools among others:
Passages for a question, with the people and projects in them and their current facts.
startold startstated insaid indecidedleadsProject AtlasStart Oct 15Start Sept 1Planning callMia RothClara WeissJonas WeberNordwerkatlas-webTobias WolfMigrations in their own PRsearch("When does Atlas start?")2 passages from the planning call and the roadmap, with Atlas, its start and its lead
In Claude Code, one command does it. You then sign in in the browser, as with "Sign in with Google":
claude mcp add --transport http --scope user scibly https://app.scibly.com/api/mcp
Other assistants add https://app.scibly.com/api/mcp as an MCP server. Which sources Scibly reads is under Integrations.
