transports-wg/meetings/2026-03-04_weekly.md
2026-03-16 16:21:15 -06:00

6.1 KiB

2026-03-04 Transports WG (Weekly)

TL;DR

The meeting, facilitated by Kurtis Van Gent and Mark Roth, focused on finalizing imminent Specification Enhancement Proposals (SEPs) and discussing new protocol extensions. The core maintainers reviewed SEP 2322 (MRTR) positively, agreeing to accept SDK breaking changes in favor of the new stateless flow. SEP 2243 (HTTP Standardization) is undergoing a final revote after consolidating header names and softening intermediary error handling. Significant debate occurred regarding the Data Layer Sessions proposal, resulting in a mandate to explicitly define targeted use cases before further design review. Andreas Schlapbach introduced a new content negotiation proposal for compact responses to support local LLMs with small context windows. Finally, the group reviewed remaining gaps for the June release.w

Detailed Summary

Process & Roadmap

  • June Release Gaps: Kurtis Van Gent identified two main gaps for the upcoming specification release: pluggable transports and caching/optimization.
  • Pluggable Transports: Mark Roth clarified that pluggable transports mostly require SDK work, not SEP changes. Minor cosmetic updates to loosen the JSON-RPC requirement might be needed. Stephen Halter emphasized the need to ensure the transport abstraction still allows for JSON-RPC to support existing proxy implementations.
  • Caching & Standard IO: Caching and optimization lack volunteers and will likely be deferred until after the June release. Standard IO updates are also firmly out of scope for the current revision.

Technical Decisions: MRTR & Tasks (SEP 2322)

  • Core Maintainer Feedback: Mark Roth reported that the core maintainers gave positive feedback on the Ephemeral Multi-Round Trip Request (MRTR) SEP. Minor schema changes were added to help SDKs discriminate between tool call responses and incomplete responses.
  • URL-Based Elicitation: The group agreed that the URL-based elicitation workflow is incompatible with the ephemeral workflow and will be restricted to the persistent (tasks) workflow.
  • Backward Compatibility: The consensus was to accept breaking changes in the SDKs to implement the new MRTR flow, deferring backward compatibility concerns unless significant user complaints arise.
  • Task Integration: Luca Chang confirmed that their task-related proposals (including "jump scares" and "task continue") are inextricably tied to MRTR and will be pursued under the Transports Working Group.

Technical Decisions: HTTP Standardization (SEP 2243)

  • Proposal Updates: Mike Kistler outlined recent changes to the SEP, which is up for a revote. Header names (tool name, resource name, prompt name) were consolidated into a single MCP name header. The payload encoding was explicitly specified as the Base64 encoding of the UTF-8 value.
  • Intermediary Handling: The requirement for intermediaries was softened so that if they reject a request based solely on header values (e.g., bad encoding), they can return a standard 400 HTTP error instead of a JSON-RPC error. Stephen Halter noted that a 404 should be explicitly avoided to prevent ending the session.
  • Header Limits: Following a discussion initiated by Stephen Halter about protocol limits, the group agreed to remove strict protocol-level limits on header lengths. These limits will be replaced with a design note advising application authors against promoting overly long parameters to headers.

Technical Decisions: Data Layer Sessions

  • Defining Use Cases: Mark Roth strongly emphasized that the sessions SEP cannot proceed without a concrete list of targeted use cases. Scenarios like "session per tool call" are mutually exclusive with "retaining state across tool calls".
  • Scope of State: Shaun Smith argued that sessions should manage opaque, application-layer state (e.g., associating a chat thread with a sandbox) rather than MCP protocol state. Peter Alexander challenged this, asking if the LLM model should be in control of the session state instead of the application.
  • Concurrency: Stephen Halter raised concerns about state synchronization during parallel tool calls. Shaun Smith clarified that the proposal currently uses a "last write wins" strategy to intentionally discourage high-stakes strict ordering, advising it be used for low-stakes storage.
  • Next Steps: The group agreed to hold a special session next week to review an updated document focused on use cases and design, which will inform a final recommendation to the core maintainers (including the potential recommendation to drop sessions entirely).

Proposals: Content Negotiation

  • Compact Responses: Andreas Schlapbach proposed a two-phase content negotiation extension (similar to HTTP Accept) allowing clients with small context windows (like local LLMs on mobile devices) to request "compact" responses from servers.
  • Protocol vs. Application Layer: Simon Russell and Kurtis Van Gent questioned why this negotiation needs to live at the protocol layer instead of just passing expressive parameters within individual tool calls. Andreas argued that an upfront session-level negotiation reduces chattiness and compute waste. Andreas will continue to refine the proposal and gather feedback in the repository.

Action Items

  • Shaun Smith to update the Sessions document by Friday/Saturday, focusing on defining use cases and the argument for keeping sessions at the application layer rather than the model layer.
  • Shaun Smith to check availability for a special session next Tuesday or Friday to discuss the updated session document.
  • Shaun Smith & Kurtis Van Gent to work together to ensure the Sessions SEP explicitly outlines the context for application versus transport sessions.
  • Mike Kistler to update SEP 2243 to replace the strict limits on header lengths with an implementation note/design guidance.
  • Kurtis Van Gent to follow up with Kryspin and Stephen Halter to formulate a formal proposal/plan for pluggable transports.