The takeaway
MCP now runs on a predictable six-month release train with freeze windows and a release manager, and the draft file-transfer extension pushes bytes out of model context into client-host transfer targets - both changes point to a protocol built to be planned around rather than chased.
Why it matters for builders
Treat MCP as a calendar, not a moving target: releases land twice a year with published proposal deadlines, soft freeze, SDK betas and hard freeze, so pin spec revisions and test Tier 1 SDK betas six weeks before each release. Design file handling around client-side transfer targets instead of base64 in tool arguments, and decide explicitly whether hosts omit or include MCP credentials on outbound transfers and whether to enforce an origin allowlist.
MCP Finally Gets a Release Schedule, and File Transfers
The Model Context Protocol has spent two years growing faster than its own process could absorb. This week its maintainers did something less glamorous than shipping a feature: they put the protocol on a calendar. In the same window they published the first draft of an extension that moves real files through MCP without stuffing megabytes of base64 into a model's context.
Together the two posts describe where MCP is heading into 2027: fewer surprise changes, more optional extensions, and a protocol that behaves like infrastructure instead of a side project.
What happened
On November 11 the MCP maintainers moved specification releases to a fixed six-month cadence, shipping on the last Tuesday of March and September, starting with version 2027-03-30. The change is written up as SEP-3398: Release Cycle Updates by Kurtis Van Gent, worked out with the Core Maintainers at MCPCon Europe, and the maintainers call it de facto release guidance from now on (MCP blog).
Two details matter more than the dates. A proposal that is not ready will slip to a later release rather than delay the specification, so the protocol stops being hostage to its slowest unfinished idea. And the project formalized a Release Manager role, with Luca Chang, co-author of the original Tasks proposal, joining Den Delimarsky and David Soria Parra.
The schedule behind the schedule is the interesting part. Planning starts eight months out, Core Maintainer review at six. Soft freeze lands two and a half months before release, when every accepted proposal must be Final and merged. Tier 1 SDK betas arrive at six weeks. Hard freeze hits one month out, and after that only critical security or release-blocking interoperability fixes get in, by full Core Maintainer vote. Late SDKs are documented as gaps, not used as an excuse to move the spec date.

The file transfer extension
A day earlier, Den Delimarsky and Wil St. Charles published a draft extension for file uploads and downloads, developed by a File Uploads Working Group over the past month and published for experimentation before it stabilizes, with the maintainers targeting stability before the end of the year.
The design is deliberately boring in the right places. A server marks a tool argument with x-mcp-file in its input schema, the same pattern already used for x-mcp-header, and the client fills it in rather than the model. Small files can travel inline as base64, capped by a server-advertised maxInlineBytes. Anything larger triggers a handshake: the client calls the tool with file metadata only, the server returns an upload_required result carrying a signed upload URL, headers, an expiry and a tamper-protected requestState blob, the client PUTs or POSTs the bytes straight to storage outside JSON-RPC, then retries with requestState unchanged. The server verifies the retry, confirms with storage that the bytes arrived, and only then runs the tool.
Downloads are the mirror image, with one structural difference: the tool returns download targets in _meta alongside its normal result, so no retry is needed. The model is told what was produced in plain text. The bytes and the URLs never enter its context.
That is the whole thesis. An agent saving a 40 MB PDF to cloud storage should not spend 40 MB of context, hallucinate a file handle, or decide where bytes physically go. The extension pushes file movement into the client and host layers, where controls exist, and keeps the model at the level of intent.

Why the two announcements belong together
The extension also shows how MCP is accumulating capability now: not by editing the core protocol, but by adding optional extensions that servers and clients can advertise. The file extension identifies itself as io.modelcontextprotocol/files and works with protocol revision 2026-07-28 and later. A client that does not support it ignores x-mcp-file and treats the argument as an ordinary object. Nothing breaks; the feature simply is not there.
That is the path MCP Apps took to become an official extension, and it is why a six-month release train is workable: the core spec freezes, extensions keep moving, and the ecosystem gets a stable target twice a year instead of a river of small breaking changes.
The file extension also inherits the stateless core shipped in the 2026-07-28 specification. requestState is self-contained, so a different server instance can finish the job without shared session state, and tools that must not run twice can mark the value used in a shared record. Multi Round-Trip Requests make the retry pattern legal in the first place.
None of this is free of risk, and the draft is candid about it. A client must not attach its MCP credential to every URL a server hands it, because that leaks the token to an unrelated storage service. So every target carries a credentials field: omit means the target authorizes itself, usually via pre-signed URL; include means the client attaches its MCP credential, but only to a path on its MCP address or the server's own origin. Clients must also refuse redirects and reject targets that supply their own Authorization header.
A residual hole remains and the maintainers say so: with omit, a malicious server can still point an upload at an endpoint it controls, and the host approves the transfer before knowing the destination. The draft recommends restricting omit targets to an origin allowlist and asks whether that should be mandatory, and it is equally blunt that a matching digest proves the bytes match a hash, not that the file is safe.

Builder impact
If you run agents in production, three things change your planning.
First, stop treating MCP as a moving target. Six-month releases with proposal deadlines, soft freeze, SDK betas and hard freeze give you a real upgrade calendar. Pin clients and servers to a spec revision, test Tier 1 SDK betas when they land six weeks before each release, and budget a review window instead of reacting to changelog surprises.
Second, the file path is about to become first-class plumbing. Today's workaround, base64 inside tool arguments, burns context, breaks on large payloads and hides failures inside JSON-RPC errors. If your automations move documents, exports or datasets, expect transfer targets to become the normal interface and design around clients that handle transfers instead of models that pretend to.
Third, credential scoping is now a design decision. Whether your client omits or includes credentials on targets, and whether you run an origin allowlist, determines how much damage a hostile MCP server can do. For agents wired into email, cloud storage and internal systems, it is the webhook-signing and OAuth-scope threat model arriving faster.
MCP is entering its boring phase, and boring is what infrastructure should be. A protocol with a release train, freeze windows, a release manager and optional extensions is a protocol you can put on a roadmap. The file draft is open for feedback through the working group, and the next spec cycle has a date.
What to watch
The March 30, 2027 deadline is the forcing function. Watch which proposals reach Final by the January soft freeze, whether the file extension stabilizes before year end as promised, and whether origin allowlists become mandatory in the final text. Those three signals show how much of MCP's awkward growth is genuinely behind it.
The Automation Brief
Read 5 AI stories instead of 50.
The essential moves in AI agents, models, automation and infrastructure — filtered for builders and operators, with the part that actually matters.
No noise. Unsubscribe anytime.
Editorial notes
Stefan Trbojevic
n8n Lab Editorial
11 November 2026
11 November 2026
Sources
AI disclosure: AI assisted with research and drafting. Factual claims are reviewed by an editor.




