Python — Discover Capabilities#

List the tools a node can see on the mesh with list_tools:

import time
from net_sdk import list_tools
 
# `node` is a MeshNode handshaked with a peer that served tools.
deadline = time.monotonic() + 3.0
while time.monotonic() < deadline and len(list_tools(node)) < 1:
    time.sleep(0.02)          # folding is asynchronous — poll until it appears
 
for t in list_tools(node):
    print(t.tool_id, "v" + str(t.version), "tags=", t.tags)

Folding is asynchronous — an announcement takes a moment to propagate, so poll until the tool you expect appears rather than assuming it's there on the first call. Announcements propagate multi-hop (bounded by a hop count), so a tool can come from a node several hops away.

Tool descriptors lower to provider tool-call formats via the openai helpers, so a discovered tool feeds straight into a chat-completion request.

Filtering nodes by capability#

Capability-filter node discovery (the find_nodes surface, e.g. "GPU nodes with ≥24 GB VRAM") is available through the node's native handle in Python rather than a clean MeshNode method. The filter model is identical across bindings — see Capabilities for the predicate surface and the CLI equivalent (net cap query --tag …). For most Python agent code, the tool API above is the path you want.

Discovery is advisory — it tells you who can, with no exclusivity.