alonf/mcppythondemo/mcp-roots-snapshot-paging

MCP Roots + Snapshot Paging Skill

Use When

You need to add a safe sandboxed read surface to an MCP server without breaking older tools.

Pattern

1. Keep the new sandbox feature additive:

• preserve older tools unchanged
• add a dedicated request_*_access tool plus a dedicated snapshot tool

1. Put all normalization and authorization in one validator:

• absolute-path requirement
• traversal rejection
• supported-root restriction
• symlink-escape rejection
• allowed-root match before reading

1. Reuse the existing snapshot/resource UX:

• tool returns {resource_uri, paginated_resource_template}
• resource returns metadata + entries + pagination

1. Page different source kinds with the same envelope:

• file snapshots → line entries
• directory snapshots → deterministic child metadata entries

1. Tell the model the order explicitly:

• request_*_access first for blocked paths
• create_*_snapshot second
• read_resource third

1. Be defensive with MCP resource query handling:

• register the paged template
• also parse ?limit=...&offset=... if it leaks through the base {id} route

Reference Files

• src/mcp_linux_diag_server/tools/proc_snapshots.py
• src/mcp_linux_diag_server/server.py
• src/mcp_linux_diag_server/client.py
• tests/test_proc_snapshots.py
• tests/test_m7_http.py