github.com/evalstate
Connecting Context: Future Transports
MCP Connect, Paris
Shaun Smith
February 2026
github.com/evalstate

Shaun Smith @evalstate

  • Open Source @ Hugging Face

  • MCP Maintainer / Transports WG

  • huggingface/hf-mcp-server

  • huggingface/upskill

  • huggingface/skills

  • huggingface.co/datasets/mcp-clients

  • Maintainer of fast-agent


github.com/evalstate

github.com/evalstate

HF Hub MCP Server API Activity (from Jun 2025)

github.com/evalstate

What's going on??

Initialize Requests?

  • 1% MCP Traffic -> Tool Call

  • Unreliable proxy for MCP install

Tool Calls: More != Better

  • Human vs. Agent usage... high call rate may indicate failure.

Sessions with > 1 Tool Call

  • 2% Conversion rate from Initialize

github.com/evalstate

Last 6 weeks of activity

github.com/evalstate

Open Source Client Data (HF Hub MCP)!!

https://huggingface.co/datasets/evalstate/mcp-clients

github.com/evalstate

Preparing MCP for the future

github.com/evalstate

Preparing MCP for the future

github.com/evalstate

Change 1: Stateless Protocol

github.com/evalstate
Current Protocol - Stateful Connection
github.com/evalstate
Future Protocol - No Initialize, Capabilities with Request/Response Pair
github.com/evalstate

Change 2: - MRTR (Multi Round Trip Requests)

github.com/evalstate
Sampling and Elicitation
github.com/evalstate
Multi-round trip flow + stateful request
github.com/evalstate
Chat interaction vs API payload
github.com/evalstate
MRTR flow + Cumulative Request
github.com/evalstate

Change 3: Real Sessions (Cookies)

github.com/evalstate
MCP cookies for session semantics
github.com/evalstate

Supporting Changes

Server Cards for advertising capabilities

Including MCP data layer content within HTTP Headers.

Clearer definition of Sampling / Elicitation usage:

w450

github.com/evalstate

Transport WG / Relevant SEPs

SEP-1442 - Make MCP Stateless by Default

SEP-XXXX: Multi Round-Trip Requests

SEP-XXXX: Notice to clarify Sampling and Elicitation Request behaviour

SEP-XXXX: Notice to clarify Sampling and Elicitation Request behaviour

github.com/evalstate

Thanks to the Transport Working Group

Initialize Events might tell us whether someone has the MCP Server installed

Raw Tool Calls is a potentially misleading vanity metric

Sessions that convert to at least one tool call

typical ratio is 1.77% of initialize events are "interesting"

## Robust Agent Loops of hundreds of turns ## Adoption of MCP Apps at Internet Scale ## Create, control and manage Resources

discoverable with a "discovery" endpoint