diff --git a/src/collections/blog/2026/04-14-mcp-in-2026-durable-standard-or-transitional-layer/hero-image.svg b/src/collections/blog/2026/04-14-mcp-in-2026-durable-standard-or-transitional-layer/hero-image.svg new file mode 100644 index 0000000000000..e9b1c284ae2c3 --- /dev/null +++ b/src/collections/blog/2026/04-14-mcp-in-2026-durable-standard-or-transitional-layer/hero-image.svg @@ -0,0 +1,77 @@ + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + + CLI + + + MCP + + + + + + + + + MCP in 2026 + + + Durable Standard or Transitional Layer? + + + + + + + diff --git a/src/collections/blog/2026/04-14-mcp-in-2026-durable-standard-or-transitional-layer/index.mdx b/src/collections/blog/2026/04-14-mcp-in-2026-durable-standard-or-transitional-layer/index.mdx new file mode 100644 index 0000000000000..bc432889d3902 --- /dev/null +++ b/src/collections/blog/2026/04-14-mcp-in-2026-durable-standard-or-transitional-layer/index.mdx @@ -0,0 +1,150 @@ +--- +title: "MCP in 2026: Durable Standard or Transitional Layer?" +subtitle: "The Model Context Protocol tripled in adoption last year. The hard questions are just starting." +date: 2026-04-14 10:00:00 -0500 +author: Lee Calcote +thumbnail: ./hero-image.svg +darkthumbnail: ./hero-image.svg +description: "MCP adoption exploded in 2025-2026, but benchmarks show CLIs outperforming MCP for agent tasks. We examine where MCP wins, where it stumbles, and whether the protocol is here to stay." +type: Blog +category: AI +tags: + - ai + - mesheryctl + - Meshery + - Engineering + - Platform Engineering +resource: true +published: true +--- + +import { BlogWrapper } from "../../Blog.style.js"; +import { Link } from "gatsby"; +import Blockquote from "../../../../reusecore/Blockquote"; +import Callout from "../../../../reusecore/Callout"; +import CTA_FullWidth from "../../../../components/Call-To-Actions/CTA_FullWidth"; +import CTAImg from "../../../../assets/images/meshery/icon-only/meshery-logo-shadow.webp"; + + + +
+

+ In the first post of this series, I argued that CLIs are not APIs - they're human interfaces that, when designed well, enable agents as a natural consequence. That post ended with a question: where does the Model Context Protocol fit? MCP adoption has been explosive. The number of MCP servers tripled in six months. OpenAI, Google DeepMind, and the major framework vendors all adopted it. But adoption and durability are different things. The hard questions about MCP are just now arriving. +

+
+ +## The Rise: What Happened + +Anthropic launched MCP in November 2024 as an open standard for AI-to-tool integration. The premise was clean: instead of writing custom integration code every time you want an AI agent to interact with an external service, standardize the communication layer. One protocol. Universal tool discovery. Structured input/output contracts. + +The industry responded fast. By early 2025, OpenAI adopted MCP across the Agents SDK, Responses API, and ChatGPT desktop. Hugging Face, LangChain, and Deepset integrated it into their frameworks. By August 2025, there were 425 registered MCP servers. By February 2026, that number hit 1,412 - a 232% increase in six months. Industry analysts project 75% of API gateway vendors will add MCP features by end of 2026. + +This is not a niche protocol. This is a standard with real momentum. + + +

MCP servers grew from 425 (August 2025) to 1,412 (February 2026). But only 27% of servers were updated in the last 30 days, and 29% haven't been touched in six months or more. Growth is real. So is abandonment.

+ + +## Where MCP Wins - and CLI Can't Touch It + +MCP solves problems that CLIs were never designed to solve. Recognizing this matters, because the "MCP vs CLI" framing that dominates the current discourse obscures the fact that they address fundamentally different concerns. + +**Enterprise governance.** MCP provides OAuth-based authentication, per-user permissions, and structured audit trails out of the box. When an agent uses an MCP server to create a Jira ticket, the action is authenticated against a specific user identity, scoped to their permissions, and logged. A CLI invocation in a shell has none of this natively. You can bolt it on - service accounts, API keys passed as environment variables, wrapper scripts that log to an audit system - but you're rebuilding what MCP provides as a first-class primitive. + +**Stateful multi-step workflows.** Some agent tasks require persistent state across multiple tool invocations. Browser automation is the canonical example - you need a persistent browser session, page context, element references that survive between calls. MCP servers maintain this state. A CLI invocation is stateless by design: it runs, it exits, its process dies. Piping state between CLI calls is possible but brittle. MCP handles it natively. + +**Sandboxed environments.** Not every agent has a terminal. Browser-based agents, chatbots, mobile assistants, sandboxed enterprise environments where shell access is locked down - MCP is the only interface available. CLI simply doesn't apply. + +**Multi-agent orchestration.** When ten agents from different teams need to coordinate across shared services - each with distinct permissions, each producing audit logs, each respecting rate limits - MCP's governance model is the right abstraction. A shared shell with ten agents writing to stdout is a disaster. + +## Where CLI Wins - and MCP Stumbles + +The benchmarks tell a clear story for developer-facing workflows. + +Comparative studies of identical tasks run through both MCP and CLI interfaces show CLI achieving significantly higher task completion rates with roughly equivalent token budgets. The tokens aren't saved - they're spent more effectively. Instead of parsing tool descriptions and navigating protocol overhead, the agent spends tokens on reasoning. + +The reasons are structural. + +**Context window cost.** MCP tool descriptions are expensive. GitHub's MCP server exposes over 40,000 tokens of tool definitions. Every MCP server an agent connects to adds to this cost. Claude Code's Tool Search feature mitigates this with lazy loading - reducing context usage by up to 95% in some cases - but that's a patch over a fundamental scaling problem. The original abstraction doesn't handle large tool surfaces gracefully. + +**Agents already know terminals.** Language models were trained on billions of lines of shell transcripts, man pages, and CLI documentation. They know how to construct `git commit -m "message"` and `kubectl get pods -n production -o json`. They don't have the same depth of training data for MCP tool invocation patterns. CLI fluency is baked into the weights. MCP fluency has to be taught. + +**Unix composition.** CLI output pipes to `jq`, `grep`, `awk`, `sort`. Decades of operational tooling snaps into place. An agent that runs `mesheryctl model list --output ndjson | jq '.name'` is leveraging an ecosystem that MCP can't replicate. MCP responses live inside the protocol boundary - they don't compose with the broader environment. + +**Debuggability.** When a CLI command fails, you see stderr. You see the exit code. You can rerun it with `--verbose`. When an MCP interaction fails, you're inside an abstraction layer that obscures the underlying cause. For developer workflows, this opacity is a real cost. + +
+ +## The Security Problem No One Solved Yet + +MCP prioritized simplicity and interoperability over security. The protocol's architects made this tradeoff deliberately - get adoption first, harden later. The consequences are arriving on schedule. + +In Q3 2025, researchers spotted the first malicious MCP server in the wild. The attack vectors are documented and growing: tool poisoning (malicious descriptions that manipulate agent behavior), silent tool redefinition (a server changing its tool schemas between calls), cross-server tool shadowing (a malicious server intercepting calls meant for a trusted one), and prompt injection through tool responses. + +CLI tools aren't immune to security issues. But they benefit from decades of hardening - sandboxed execution, well-understood permission models, process isolation, established patterns for credential management. MCP's security model is still being written. For enterprise teams evaluating MCP for production workloads, this gap is not academic. + + +

MCP's security model is evolving. If you're deploying MCP servers in production, implement your own authentication layer, audit logging, and input validation. Don't assume the protocol handles it.

+ + +## The Consolidation Phase + +Every technology adoption curve hits a consolidation phase. MCP is entering it now. + +The raw numbers look impressive - 1,412 servers and counting. But dig into the data and a different picture emerges. Only 27% of registered MCP servers were updated in the last 30 days. 29% haven't been touched in six months. The initial gold rush produced hundreds of quick integrations, many of which are already abandonware. + +This is normal and healthy. The signal-to-noise ratio improves as the ecosystem matures. The winners - Chrome DevTools MCP (15,000 GitHub stars in three months), Playwright, Supabase, the major cloud provider integrations - are pulling away from the pack. + +But consolidation also means the early promise of "one protocol to connect everything" is meeting reality. The MCP servers that survive will be the ones backed by organizations with the resources to maintain them, keep security patches current, and evolve their tool schemas alongside their products. + +## A Decision Framework + +The "MCP vs CLI" framing is a false binary. Here's how to think about it instead. + +**Use CLI when** the agent operates in a developer environment with terminal access, the task is stateless or can be composed from stateless steps, token efficiency matters, you need Unix-style composition with existing tools, and the target audience is engineers who debug in the terminal. + +**Use MCP when** the agent operates in a sandboxed or browser-based environment, the task requires persistent state across multiple invocations, enterprise governance is a hard requirement (per-user auth, audit trails, rate limiting), multiple agents need coordinated access to shared services, and the target audience is non-technical users interacting through a chat interface. + +**Use both when** you need the execution efficiency of CLI with the governance properties of MCP. This is Meshery's approach: the CLI handles the work, and an MCP surface (`mesheryctl mcp`) provides the governance wrapper for enterprise and multi-agent scenarios. + +## What This Means for How You Build + +If you maintain a CLI tool - any CLI tool - the lesson from the MCP discourse is not "go build an MCP server." The lesson is: make sure your CLI is designed for scriptability first. Non-interactive flags for every prompt. Machine-readable output. Schema introspection. Strict input validation. Deterministic exit codes. + +If you do that, your CLI serves both humans and agents. And if you later need an MCP layer for governance, you build it on top of a solid CLI foundation. The MCP server calls the CLI under the hood. The protocol provides auth, audit, and orchestration. The CLI provides speed, reliability, and composability. + +If you skip the CLI and go straight to MCP, you get a governance layer with nothing efficient underneath it. Your agents spend tokens on protocol overhead instead of reasoning. Your developers can't debug with a terminal. Your automation doesn't compose with the Unix ecosystem. + +
+ +## The Honest Forecast + +MCP is not going away. The adoption is too broad, the backing too strong, the enterprise need too real. Google, OpenAI, Anthropic, and the major framework vendors are committed. The protocol will mature, the security model will harden, and the consolidation will produce a smaller set of high-quality, well-maintained servers. + +But MCP is also not the universal interface that the initial hype suggested. The "MCP for everything" era is already ending. Teams are learning - sometimes painfully - that CLI is faster, cheaper, and more debuggable for the majority of developer-facing agent workflows. MCP earns its place in enterprise governance, multi-agent coordination, and sandboxed environments. Outside those contexts, it's overhead. + +The tools that will define the next phase of agentic infrastructure are the ones that get this layering right. CLI for execution. MCP for governance. Both designed with the same first principle: build for the human, enable the agent. + + + +
+

+ The mesheryctl functional design spec is open for community input, including the MCP surface proposal. Join the conversation in the Meshery Slack or explore the Meshery playground to see agent-native CLI design in practice. +

+
+ +