Python — Move Artifacts#

The bus is a coordination layer, not a file transfer. Move large data as a content-addressed blob and put only a small reference on the bus.

Reference, don't embed#

node.emit({"frame_id": "abc123", "blob": blob_ref})   # small event carries the ref

Blobs are addressed by their BLAKE3 hash — the address is the content — and live on whichever nodes have capacity, drifting toward the nodes that read them.

Fetch and store over the mesh#

The transfer operations are functions in net_sdk.transport that take the mesh node first. Install the transfer engine once, then fetch:

from net_sdk import transport
 
transport.serve_blob_transfer(mesh, adapter)          # install once per node
data = transport.fetch_blob(mesh, holder_id, blob_ref)  # stream from a known holder
data = transport.fetch_blob_discovered(mesh, blob_ref)  # or let the mesh find a holder
transport.fetch_dir(mesh, holder_id, root_ref, dest)    # directories, materialized atomically

Peak memory is one chunk (~4 MiB) regardless of total size, and a directory fetch either becomes the complete tree or leaves the destination untouched. A producer that reads back its own write never sees a gap (read-your-own-writes).

The exact signatures, the storage/gravity model, and the operator CLI (net-mesh transfer …) are in Blob Storage (Dataforts) and the CLI Reference.

Bridged tools have no artifacts#

Artifacts are a native capability. Tools brought in through the MCP bridge are mcp_bridge tier — request/response only. If your work needs to move bytes, it's a native capability.