---

name: client-protocols description: "Build and troubleshoot Triton KServe HTTP/gRPC request payloads and client-side protocol usage." disable-model-invocation: true metadata: disco-role: operating license: NOASSERTION

Client Protocols

Use this sub-skill when the user needs Triton HTTP/gRPC request shapes, health/metadata queries, binary input encoding, shared-memory guidance, or client-library request recipes.

Route Within This Sub-skill

• KServe v2 HTTP/gRPC payload structure, health, metadata, repository-control, binary data, and status mapping: read references/kserve-http-grpc.md.
• Client-library recipes, curl vs Python client, streaming, timeouts, and shared-memory usage: read references/client-library-recipes.md.
• Transport mismatches, tensor name/shape/datatype mistakes, invalid argument responses, and binary/body encoding problems: read references/troubleshooting.md.
• Helper to build request descriptors without contacting a live server: run scripts/build_kserve_request.py.

If the user needs to author the model repository or config, route to ../model-repository-and-config/SKILL.md. If the user is launching Triton or checking metrics/ports, route to ../server-runtime-and-deployment/SKILL.md.

Safe Default Workflow

1. Confirm the target endpoint family: HTTP/REST or gRPC.
2. Identify the model name, version, tensor names, shapes, and datatypes from repository/config or metadata.
3. Build a request descriptor or curl/client example with the bundled helper before sending anything live.
4. Distinguish KServe /v2/* payloads from Triton's OpenAI-compatible /v1/* payloads.
5. Use live metadata or a ready endpoint only when the user explicitly allows contacting a server.