MCP publishes new roadmap, prioritizes HTTP and agent identity

The Core Maintainers of the Model Context Protocol (MCP), the open standard hosted by the Linux Foundation's Agentic AI Foundation (AAIF), have published an updated roadmap setting the direction for the protocol's next specification release and beyond. As Japanese tech outlet Publickey reported, the roadmap is organized around five priority areas: agentic messaging primitives, HTTP-native transport unification, agent identity and enterprise-ready security, improved tool primitives, and a better SDK developer experience. The official MCP blog confirms the roadmap was developed jointly by Core Maintainers and the community of Working Groups, and that it "sets the direction for protocol work over the coming months."

MCP began as a protocol Anthropic proposed in November 2024 to connect its Claude models to external data sources and tools. On December 9, 2025, the Linux Foundation announced it had received the donation of the protocol and formed the Agentic AI Foundation to govern its evolution, with Amazon Web Services, Anthropic, Block, Bloomberg, Cloudflare, Google, Microsoft and OpenAI as platinum members. According to the Linux Foundation's announcement, MCP became one of the fastest-growing open-source projects of the past year, with over 97 million monthly SDK downloads and more than 10,000 active servers, and first-class support in ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot and Visual Studio Code. This July, the Enterprise-Managed Authorization (EMA) extension reached stable status and the protocol moved to stateless connections, a major update that made large-scale deployments easier to run.

Five priority areas

According to the official post "The New MCP Roadmap," this update builds on the previous roadmap published in March, which had four priorities: transport evolution and scalability, agent communication, governance maturation, and enterprise readiness. Most of the progress from the past five months landed in the 2026-07-28 specification release. Protocol-level sessions and the initialization handshake were removed so servers can scale horizontally without holding state (SEP-2575, SEP-2567), clients gained a `server/discover` call to learn a server's capabilities before connecting, and list results became cacheable (SEP-2549).

• Agentic messaging primitives: build out server-initiated events (webhooks and channels) so clients stop polling for results, and mature the Tasks extension (SEP-2663) toward inclusion in the core specification
• HTTP-native transport unification: extend Streamable HTTP so it covers local servers too, using HTTP/2 over stdio to keep the multiplexing benefits of a subprocess connection
• Agent identity and enterprise-ready security: finalize Demonstrating Proof of Possession (DPoP) and define agent identity and delegation through Workload Identity Federation (SEP-1933) and ID-JAG token exchange, rather than pasted API keys or long-lived tokens
• Improved primitives: standardize the shape of `tools/call` responses and introduce progressive discovery, so a model doesn't have to ingest a server's entire tool catalog before a user asks a single question
• Improved SDK developer experience: sharpen API ergonomics and documentation accuracy, since many developers — and increasingly AI agents themselves — now write MCP client and server code by pointing at the SDKs

What this solves for enterprises

Today's MCP authorization model assumes a human approves access in a browser. The official blog notes that a growing share of callers are instead agents running as cloud workloads with their own identity, acting on behalf of an absent user, or delegating narrower authority to sub-agents. The roadmap's identity work aims to give MCP servers a standardized way to recognize and trust those agent identities, built on existing standards rather than pasted API keys, coordinating with the IETF OAuth and WIMSE working groups on the underlying token-exchange mechanics. The push to unify on HTTP matters just as much operationally: a remote MCP server built on Streamable HTTP is now no different from any other HTTP workload, so companies can host and scale it on the same infrastructure — load balancers, gateways, observability — they already use for their APIs, instead of building bespoke plumbing for a separate protocol.

On prioritization, the official blog is explicit: proposals (SEPs) that map to one of the five priority areas get expedited review, while others aren't automatically rejected but face longer timelines and a higher bar for justification, since maintainer review capacity is limited and is being pointed deliberately at these five areas.

Why this matters for teams building agents

MCP is already the connector most major AI products — ChatGPT, Claude, Cursor, Gemini, Microsoft Copilot, Visual Studio Code — use to reach outside tools, which is what makes this roadmap more than an internal engineering document: it signals where the de facto standard for agent-to-tool connectivity is headed next. For companies wiring AI agents into internal systems, the HTTP unification work should lower the operational cost of running MCP servers at scale, since they stop needing separate infrastructure for stdio or WebSocket-based deployments. The identity work addresses a very concrete risk teams are already carrying: API keys shared across services or left valid after an employee changes roles. A standardized way to scope and revoke an agent's own credentials, backed by audit trails, replaces informal key-sharing with something a security team can actually govern. Organizations that currently authenticate MCP connections with long-lived tokens should treat this as an early signal to plan a migration path once the identity and delegation work ships.