Where Model Context Protocol is heading and what it means for teams building production AI agents

MCP in 2026: from developer experiment to enterprise infrastructure

The Model Context Protocol  has crossed a threshold. What began as Anthropic’s open-source experiment in November 2024 to connect AI models with external tools is now the de facto standard for AI agent integration, with 97 million monthly SDK downloads, over 9,400 public servers, and native support from every major AI provider: Anthropic, OpenAI, Google DeepMind, and Microsoft.

In December 2025, Anthropic donated MCP to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation co-founded by Anthropic, Block, and OpenAI. The move cemented MCP’s status as a vendor-neutral open standard governed by a community process rather than a single company’s product decisions.

The 2026 roadmap, published in March by lead maintainer David Soria Parra, reflects this shift. It is no longer organized around release milestones. Instead, it defines four priority areas driven by Working Groups and Interest Groups, each focused on a specific set of production challenges that early deployments keep surfacing.

Priority 1: transport evolution and scalability

Streamable HTTP, the transport that lets MCP servers run as remote services rather than local processes, unlocked a wave of production deployments in 2025. But running it at scale has revealed consistent gaps.

Stateful sessions fight with load balancers. Horizontal scaling requires workarounds. There is no standard way for a registry or crawler to discover what a server does without establishing a live connection. The 2026 roadmap addresses these directly:

Next-generation transport. Evolve Streamable HTTP to run statelessly across multiple server instances and behave correctly behind load balancers and proxies. This is the single most requested change from teams running MCP in production.

Scalable session handling. Define how sessions are created, resumed, and migrated so that server restarts and scale-out events are transparent to connected clients. Currently, losing a session means losing context.

MCP Server Cards. A standard for exposing structured server metadata via a .well-known URL, so browsers, crawlers, and registries can discover a server’s capabilities without connecting to it. This enables the server directories and marketplaces (Smithery, Glama, PulseMCP) that are emerging to index and compare servers programmatically.

Priority 2: agent communication

MCP currently defines how a single AI model interacts with tools. But production agent architectures increasingly involve multiple agents collaborating on complex workflows: one agent qualifies leads, another drafts outreach, a third validates compliance. The 2026 roadmap extends MCP to handle these multi-agent patterns.

Tasks. Asynchronous task handling that lets agents launch work in one request and retrieve results later. This is critical for long-running operations and non-blocking workflows. Early production use has already exposed gaps around retry semantics and result lifecycle management.

Agent-to-agent communication. While Google’s A2A (Agent-to-Agent) protocol handles inter-agent coordination at a higher level, MCP’s roadmap focuses on the tool-level primitives that enable agents to delegate and receive results within the MCP framework.

Streaming and progressive results. Support for agents that need to process large outputs incrementally rather than waiting for a complete response. Essential for data-heavy workflows where results arrive over time.

This priority is particularly relevant for teams building AI agent architectures with multiple specialized components that need to coordinate through shared tool access.

Priority 3: governance maturation

MCP has grown into a multi-company open standard under the Linux Foundation. The formal governance process introduced in 2025 through Spec Enhancement Proposals (SEPs) and Working Groups established the foundation. The 2026 roadmap builds on it:

Contributor ladder. A formal progression from community participant to Working Group contributor to WG facilitator to lead maintainer to core maintainer, with explicit nomination and review criteria at each step. This ensures the project doesn’t depend on a small set of individuals.

Delegation model. Working Groups with a proven track record can accept SEPs and publish extension updates within their domain without a full core-maintainer review cycle. This speeds up development in mature areas.

Charter template. Every Working Group and Interest Group maintains a public charter: scope, active deliverables, success criteria, and retirement conditions, reviewed quarterly.

The governance priority reflects a practical reality: MCP now has more community contributors and proposals than a small core team can review. The protocol needs to distribute decision-making to scale its own development.

Priority 4: enterprise readiness

This is the least defined of the four priorities, and the roadmap says that’s intentional. The maintainers want the people experiencing enterprise deployment challenges to help define the work.

The problems are well-documented. Teams deploying MCP at enterprise scale keep running into the same blockers:

Audit trails. No standardized way to log which tools were called, by which agent, with what arguments, and what results were returned. Regulated industries need this for compliance.

Authentication and identity. MCP currently relies on static secrets for most server authentication. The roadmap points toward OAuth 2.1 with PKCE for browser-based agents and SAML/OIDC integration for enterprise identity providers. Active proposals also cover DPoP and Workload Identity Federation.

Gateway behavior. No standard for how MCP traffic should flow through enterprise network infrastructure, proxies, and API gateways.

Configuration portability. Server configurations that work in one client don’t automatically transfer to another. There’s no standard for packaging and distributing MCP configurations across teams.

Most enterprise readiness work is expected to land as extensions rather than core spec changes, a deliberate choice to avoid making the base protocol heavier for everyone else. A dedicated Enterprise Working Group has not yet formed, and the roadmap explicitly invites enterprise infrastructure practitioners to lead it.

Connect your AI agent to human experts

Tendem MCP gives your agent access to 10,000+ vetted domain specialists on demand. One install, no code changes, non-blocking async execution.

Learn more about Tendem MCP →

The adoption landscape: where MCP stands today

The numbers tell a clear story. MCP’s TypeScript and Python SDKs reached 97 million monthly downloads in March 2026, up from approximately 2 million at launch. The growth rate of 4,750% in 16 months mirrors the adoption curves of foundational infrastructure protocols. For context, the React npm package took approximately three years to reach comparable download volumes.

The server ecosystem has grown from a handful of reference implementations to over 9,400 public servers covering databases, CRMs, cloud providers, productivity tools, developer tools, e-commerce platforms, analytics services, and specialized domains. Private and enterprise-internal servers are conservatively estimated at 3-4x that number.

Gartner projects that by the end of 2026, 40% of enterprise applications will include task-specific AI agents, and 75% of API gateway vendors will have MCP features. According to their CIO survey, 80% of enterprise applications shipped in Q1 2026 embed at least one AI agent, up from 33% in 2024. Yet only 17% of organizations have fully deployed AI agents, with more than 60% expecting to do so within two years, the most aggressive adoption curve among all emerging technologies in the survey.

The gap between adoption intent and production deployment, estimated at 68 percentage points, is the defining challenge. The organizations closing it fastest share a common trait: they treat evaluation, governance, and human oversight as infrastructure requirements, not afterthoughts.

What’s on the horizon

Beyond the four core priorities, the roadmap identifies several areas in active development that could land if Working Groups drive them forward:

Triggers and events. MCP currently operates on a request-response model. Triggers would allow servers to push notifications to clients when external events occur, a bookkeeping system notifying the agent when a transaction is flagged, for example, rather than requiring the agent to poll.

Skills and capabilities. A higher-level abstraction where servers can advertise composable "skills" (multi-tool workflows) rather than individual tools. This would let agents reason about capabilities at a higher level of abstraction.

MCP registry. A curated, verified server directory with security audits, usage statistics, and SLA commitments, planned for Q4 2026. This would address the current fragmentation across multiple community directories.

Deeper security and authorization. Active proposals for advanced security patterns including DPoP (Demonstrating Proof-of-Possession) and Workload Identity Federation for machine-to-machine authentication.

SDK v2 and progressive discovery. Next-generation SDKs that support smarter clients capable of negotiating capabilities and adapting to different server versions dynamically.

The human expertise layer: what the roadmap doesn’t cover

The 2026 roadmap is comprehensive for the protocol itself, but one dimension it does not address directly is how agents should handle tasks that exceed their autonomous capabilities. The roadmap assumes tools are software endpoints. It does not prescribe what happens when the right response requires human judgment.

This is the gap that Tendem by Toloka fills. As MCP becomes more capable, with better transport, task handling, and enterprise auth, the agents built on it will take on more complex and higher-stakes workflows. The need for a human reliability layer does not decrease as the protocol matures; it increases. More capable agents operating in more sensitive environments make the consequences of errors larger, not smaller.

Tendem provides this layer as a standard MCP server: one install gives your agent access to 10,000+ verified domain experts who handle fact verification, research, analysis, and quality assurance. The agent delegates via a normal MCP tool call and receives structured results with verified data, sources, and quality scores. In benchmarks across 94 real-world tasks, the hybrid approach achieved 1.8x higher quality than AI-only execution. For teams that also need to understand how their agents behave inside MCP environments, Toloka’s MCP evaluations provide trajectory-based diagnostic reports with a 12-type failure taxonomy.

What this means for teams building agents today

The roadmap is not a release schedule. It is a set of priority signals that tell you where the community is investing effort and where the specification will evolve. For teams making architecture decisions now, several practical implications stand out:

Design for remote servers from the start. The transport evolution priority makes it clear that stdio (local subprocess) is the past and Streamable HTTP (remote service) is the future. Architect your server deployments for remote access now, even if you’re prototyping locally.

Build your own audit logging. Enterprise audit trails are coming to the spec, but they are not here yet. Implement structured logging of tool calls, arguments, and results in your infrastructure now. When the standard arrives, you’ll migrate rather than retrofit.

Don’t hard-code authentication. OAuth 2.1 and enterprise identity integration are on the roadmap. Design your auth flows to be replaceable so you can adopt the standard when it lands.

Plan for multi-agent coordination. If you’re building single-agent workflows today, consider how they might evolve into multi-agent architectures. The tasks and agent communication primitives on the roadmap will make this easier, but your data models and state management need to support it.

Invest in evaluation now. The 40% project failure rate Gartner warns about is largely driven by insufficient evaluation and governance. Teams running trajectory-based MCP evaluations during development catch regressions early and ship more reliable agents.

Add human expertise for high-stakes workflows. The protocol will keep improving, but no amount of transport scalability or enterprise auth fixes the fundamental limitation that AI agents lack human judgment. Build human-in-the-loop into your architecture from the beginning.

See how your agent really performs

Toloka’s MCP evaluations reveal where AI agents fail in real workflows, with trajectory analysis and human-annotated failure reports.

Talk to us →

Frequently asked questions

What are the four priorities in the 2026 MCP roadmap?

Who governs MCP now?

Will MCP replace traditional APIs?

How does MCP relate to Google’s A2A protocol?

What is the biggest risk for teams adopting MCP in 2026?

Related reading

What is Model Context Protocol (MCP)?

The importance of MCP evaluations in agentic AI

AI agents: components and their role in autonomous decision-making

From human intelligence to hybrid AI: the next chapter for Toloka