Skip to main content

Execution Boundary

How stdio servers run with Bloque​

In a typical MCP setup, stdio-based servers run as child processes on the same machine as the AI agent. They can read and write local files, access local databases, and call local services directly.

With Bloque, stdio-based servers run inside Bloque's MCP Proxy infrastructure — not on your local machine.

This means:

  • Stdio-based MCP server processes have no access to your local filesystem.
  • They cannot read files on your computer, even if a tool appears to accept a local path.
  • Network access from stdio servers goes through Bloque's infrastructure, not your local network.

Remote MCP servers (SSE/HTTP) are unaffected — they already run on remote infrastructure by nature.

The /shared filesystem​

Bloque exposes a shared filesystem at /shared that is accessible to all MCP server processes running in your Hub session. You can use this as a bridge to pass files between your local machine and Bloque-hosted stdio tools.

Uploading a file to Bloque​

  1. Call the bloque_get_upload_url tool to get a pre-signed HTTP upload URL.
  2. Use HTTP PUT to upload your file to that URL.
  3. The file is now available under /shared for any MCP server in your session.

Downloading a file from Bloque​

  1. Call bloque_get_download_url with the path under /shared.
  2. Use HTTP GET to download the file to your local machine.

Listing files​

Use standard filesystem tools (e.g. list_directory, read_file) to inspect files and directories under /shared.

Example prompt to your AI agent​

When starting a task that involves local files, tell your AI agent about this boundary:

Important context: Bloque is an MCP gateway. Stdio-based MCP servers run in
Bloque's infrastructure, not on your machine. Those tools can only access files
under /shared. To move files across the boundary:
- Upload local → Bloque: bloque_get_upload_url + HTTP PUT
- Download Bloque → local: bloque_get_download_url + HTTP GET
- Browse /shared: use filesystem tools

Providing this context upfront prevents the agent from attempting to reference local paths that the MCP server process cannot reach.