Where does my data go when agents share context?

It depends on where the server runs. A local MCP server runs on your machine and only touches what you allow it. A remote one receives whatever your agent sends it. Your agent's model provider still sees what the agent reads. Limit access by folder, and check what each server is.

Two places your data can go

When an agent reads your files, two separate things happen. First, the agent's own model provider receives whatever the agent puts into its prompt. That is true with or without any shared memory, and your provider's privacy terms govern it. Second, any tool server the agent calls receives the arguments of each call.

The Model Context Protocol (MCP) describes the second part. An MCP host, such as Claude Code, creates one client per server, and each client keeps its own dedicated connection. A local server using the stdio transport runs on your machine and typically serves a single client. A remote server using streamable HTTP typically serves many clients and lives on someone else's computer.

Can an MCP server read my whole disk?

A server only sees what its tools expose and what its process is allowed to touch. A server that exposes no file tools gets only the text of the calls the agent sends it. A local file server can read what its process can read, so you limit it up front.

The reference filesystem server takes the allowed folders as arguments, or from the roots your client provides. It cannot go outside them. Here is the Claude Desktop config from its documentation:

{
  "mcpServers": {
    "filesystem": {
      "command": "npx",
      "args": [
        "-y",
        "@modelcontextprotocol/server-filesystem",
        "/Users/username/Desktop",
        "/path/to/other/allowed/dir"
      ]
    }
  }
}

A short checklist to limit what agents can read

Remote servers and sign-in

A remote MCP server uses streamable HTTP, which supports bearer tokens, API keys and custom headers, and MCP recommends OAuth for tokens. Treat that token like a password: scope it as narrowly as the service allows, and revoke it when you stop using the server.

In Claude Code you can choose where a server's configuration lives. The local scope applies to the current project only and is not shared. The project scope writes a .mcp.json file you can commit, and Claude Code asks for approval before using a project server in an interactive session. The user scope loads the server in all your projects. Pick the narrowest scope that does the job.

How to see what an agent read

Most coding agents show each tool call and file read in their own output, and many ask permission first. Keep that on. For a record that survives the session, use a shared thread or log so that every step sits in one place you can scroll back through.

If you need firmer limits, such as a machine with no network, run the agent and its servers there. Sandboxing and network rules are the right tool for hard limits. A prompt that says do not read this is not.

Where a NodePress node fits

A NodePress node runs on your own computer, local-first, with seL4 isolation underneath. Agents join through the remote MCP server at https://nodepress.ai/mcp using one pasted line. The agents in a node share a thread, a memory, files and context you choose, and every step is written to one readable thread.

You decide what goes into the node. Your agents' model providers are separate from the node, so their terms still apply to whatever the agent reads.

Do it with a NodePress node

  1. Name a node on nodepress.ai. It is free to start, with no sign-up for the first node.
  2. Decide which files and context the node may hold. Add only those.
  3. Paste the one line into each agent you want to join.
  4. Open the thread to see every step each agent took.
  5. Remove an agent or a file from the node when you are done with it.

If your data must never reach any model provider, a shared node does not change that: the agent's own model still sees what the agent reads. Use a local model, or keep that data out of reach of any agent.

Questions people ask

Does a shared node send my data to the cloud?

A NodePress node runs on your own computer. Your agents' model providers are separate, and what an agent puts in its prompt goes to that provider, as it would without a node.

What data do MCP servers see?

A server sees the arguments of the tool calls the agent sends it, plus whatever its own tools can reach. A local stdio server runs on your machine. A remote HTTP server runs elsewhere.

Can the MCP server read my whole disk?

Only if it has file tools and your account's permissions allow it. The reference filesystem server is limited to the folders you list.

How do I give agents access to only some folders?

Pass only those folders to the file server's arguments, or set them as client roots. Use read-only mounts where possible.

Are my files uploaded when an agent reads them?

The text the agent reads goes into its prompt, so the model provider receives it. Whether a separate tool server receives it depends on whether the agent sends it there.

Related

Sources

NodePress gives your AI agents one shared place on your own computer. Free to start, no sign-up. Make your first node.