Claude is good at reasoning and bad at knowing anything about your company, because by default it knows nothing about it. Ask it about your billing plan or last week’s decision and it either guesses or asks you to paste the context in. The Model Context Protocol (MCP) fixes that by letting Claude read an external source directly. Suda is one such source: it serves your whole context graph over MCP so Claude answers from your real, current context. Here is how to set it up.

What MCP does for Claude

MCP is an open protocol that lets Claude call external servers in a consistent way. Once Claude speaks MCP, it can reach any server that exposes it: a database, a ticketing system, or a context graph that spans all your tools at once. You connect the server, and Claude gains a new place to read from without you pasting anything.

The important part is what sits on the other end. Point Claude at a single-tool server and it gets records from that one tool. Point it at a context server and it gets the connected, reconciled context across everything you use.

The manual way, and why it fails

Most people give Claude context by pasting. You copy the Notion page, the Slack thread, and the relevant ticket into the prompt and ask your question. It works once. Then the doc changes, the thread grows, and your pasted snapshot is wrong. You are also doing the retrieval by hand, which is the job you wanted Claude to do.

Pasting also caps how much context Claude can hold. Dump three long documents in and you spend the window on raw text, most of which is not relevant to the question. The answer gets worse, not better.

The MCP way

Instead of pasting, you connect a context server once and let Claude query it per question. With Suda that path looks like this:

Connect your sources. Suda ingests from more than 700, including Notion, Slack, and Linear. It builds a context graph that models how the facts connect and keeps itself current, so stale facts retire and conflicts resolve on their own.
Point Claude at the server over MCP. Claude now has a tool it can call to read your context.
Ask normally. When you ask about billing, Claude queries the graph, gets the current answer assembled across your tools, and responds. No pasting, no stale snapshot.

npx suda connect --agent claude

That command connects your first source and wires Claude to read the graph over MCP. Add the next source and it joins the same graph.

Why this beats pasting more text

Sending Claude more text is not the same as giving it better context. The graph sends only what the question needs, already reconciled, so Claude receives far less text and a clearer signal. For Suda that is about 85% less per answer, which lowers token cost as a byproduct of being correct. The window stays free for reasoning instead of raw dump.

Access stays permissioned, so Claude reads only what the person asking is allowed to see. A support agent and an engineer asking the same question get answers scoped to their access, not the whole company’s data.

What changes in practice

Once Claude reads your context over MCP, the back-and-forth disappears. You stop prefacing every question with three paragraphs of setup. A new hire’s Claude can trace a feature from its spec to the decision that changed it without anyone assembling the trail. And because the graph is shared, every agent that speaks MCP reads the same current context, not a private copy that drifts.

Start here

Connect your first source and point Claude at it:

npx suda connect

For the architecture behind the server, read MCP context servers. To connect the specific tools your team runs, see connect Notion, Slack, and Linear to your AI.