modelcontextprotocol/experimental-ext-variants

Watch

Kurt Degiorgio e1c17e047f

Add example servers for SEP-2053 variant patterns (#4 )

Add four new example servers demonstrating different SEP-2053 variant
patterns:

- model-optimized (same tools, different descriptions per LLM family)
- trading (API versioning with deprecation lifecycle)
- research (context budget management)
- Refactor the existing variants-http example into a GitHub-style server
with custom ranking.

Signed-off-by: Kurt Degiorgio <kdegiorgio@bloomberg.net>

2026-03-01 12:42:43 +05:30

3.1 KiB

Raw Blame History

Model-Optimized Weather Service

A variant-aware MCP server that exposes the same tools (get_weather, get_forecast) with different descriptions optimized for different LLM families. This is the primary use case described in SEP-2053.

Pattern demonstrated: Same tools, different descriptions per variant.

Variants

Variant	Description Style	Target
`claude-optimized`	Verbose, structured with explicit usage guidance	Anthropic Claude models
`gpt-optimized`	Concise, function-style with JSON Schema emphasis	OpenAI GPT models
`compact`	Minimal, bare-minimum descriptions	Context-constrained environments

All three variants expose the same two tools with the same behavior — only the tool descriptions differ.

Custom Ranking

Clients send a "modelFamily" hint (e.g., "anthropic", "openai"). The ranking function matches the hint to the variant's modelFamily, falling back to priority order.

Run

go run ./examples/server/model-optimized

The server listens on http://localhost:8080.

Connect any MCP client to http://localhost:8080.

Try It

1. Initialize without hints (default ranking)

curl -s -X POST http://localhost:8080/mcp \
  -H 'Content-Type: application/json' \
  -d '{
    "jsonrpc": "2.0", "id": 1,
    "method": "initialize",
    "params": {
      "protocolVersion": "2025-03-26",
      "clientInfo": {"name": "test", "version": "1.0"},
      "capabilities": {}
    }
  }'

The first variant in the response is the default — clients that don't specify a variant in subsequent requests will use it. Without hints, claude-optimized is the default:

{
  "availableVariants": [
    {
      "id": "claude-optimized",
      "hints": { "contextSize": "verbose", "modelFamily": "anthropic" },
      "status": "stable"
    },
    {
      "id": "gpt-optimized",
      "hints": { "contextSize": "standard", "modelFamily": "openai" },
      "status": "stable"
    },
    {
      "id": "compact",
      "hints": { "contextSize": "compact", "modelFamily": "any" },
      "status": "stable"
    }
  ]
}

2. Initialize with an OpenAI hint (ranking changes)

curl -s -X POST http://localhost:8080/mcp \
  -H 'Content-Type: application/json' \
  -d '{
    "jsonrpc": "2.0", "id": 1,
    "method": "initialize",
    "params": {
      "protocolVersion": "2025-03-26",
      "clientInfo": {"name": "test", "version": "1.0"},
      "capabilities": {
        "experimental": {
          "io.modelcontextprotocol/server-variants": {
            "variantHints": {
              "hints": {"modelFamily": "openai"}
            }
          }
        }
      }
    }
  }'

Now gpt-optimized is ranked first and becomes the default:

{
  "availableVariants": [
    {
      "id": "gpt-optimized",
      "hints": { "contextSize": "standard", "modelFamily": "openai" },
      "status": "stable"
    },
    {
      "id": "claude-optimized",
      "hints": { "contextSize": "verbose", "modelFamily": "anthropic" },
      "status": "stable"
    },
    {
      "id": "compact",
      "hints": { "contextSize": "compact", "modelFamily": "any" },
      "status": "stable"
    }
  ]
}

3.1 KiB Raw Blame History