Python — Announce a Capability#

In Python the ergonomic way to put a discoverable, callable capability on the mesh is the tool API. serve_tool registers a handler on your node and announces it; peers fold the announcement and can then list and call it.

from net_sdk import MeshNode, serve_tool
 
node = MeshNode(bind_addr="127.0.0.1:0", psk="42" * 32)
 
def web_search(req):
    return {"results": [f"first hit for '{req['query']}'"]}
 
handle = serve_tool(
    node,
    {"name": "web_search", "description": "Search the web.", "tags": ["web", "research"]},
    web_search,
)
# handle stays alive while the tool is served; close it to withdraw.

serve_tool / serve_tool_async (and the streaming variants) are top-level functions that take the node first — the same node-first convention as the transfer functions. The exact keyword shape of the tool descriptor is in the tool_calling example; the essentials are a name, a description, and tags.

Raw capability tags (the handle)#

If you need to announce plain capability tags (hardware, region, arbitrary tags) rather than a callable tool, that surface is reached through the node's native handle in Python — it is not a clean MeshNode method the way it is in Rust/TS. The tag/axis model itself is identical across bindings — see Capabilities and Capability Schema — but in Python prefer the tool API above unless you specifically need raw tags.

Policy: announced ≠ invocable#

Announcing makes a capability discoverable, not open. Visibility and invocation are separate; the boundary is enforced on invoke, not announce — see Invoke.