Skip to the content.

2026 (rolling) — Model Context Protocol (MCP)

Why this matters

MCP became the cross-vendor connector standard in 2025–2026. Anthropic originated it; by December 2025 the protocol was shipping 97 million monthly SDK downloads, 10,000+ active servers in production, and hundreds of distinct AI clients using it. In December 2025 Anthropic donated MCP to the newly formed Agentic AI Foundation under the Linux Foundation, backed by Block, OpenAI, AWS, Google, Microsoft, and others — MCP is no longer one company’s project.

For an agent-notes set, MCP matters because it’s the protocol under everything: when Claude Code talks to a database, a file system, an issue tracker, a SaaS product, or another agent — that conversation almost certainly happens over MCP. Skills declare MCP servers they need. Hermes Agent’s ACP work (see the Hermes entry) is partly MCP made agent-to-agent. The Anthropic finance agents ship with “connectors” that are MCP servers under another name.

This entry exists because MCP was conspicuously underexplained across the other entries — referenced in passing repeatedly, never given its own room.

What MCP actually is

A client–server protocol for connecting LLMs to external context and tools. The pieces:

In practice you encounter MCP through configuration. A Claude Code project’s .mcp.json (or the plugin’s plugin.json) declares which MCP servers to spin up; Claude Code starts the processes at session start; the servers’ tools become available to the model alongside built-in tools.

The 2026 inflection points

Governance — the Linux Foundation donation

In December 2025 Anthropic donated MCP to the Agentic AI Foundation (AAIF), a directed fund under the Linux Foundation. Founding members: Block, OpenAI, AWS, Google, Microsoft, Anthropic itself, and others. The same donation event covered AGENTS.md (see that entry).

This isn’t symbolic; it changes the protocol’s evolution path. Working Groups now drive spec updates, with cross-vendor representation. The 2026 roadmap was reorganised around priority areas rather than release dates — Working/Interest Groups own timelines for their deliverables. That’s structurally different from “Anthropic ships a new spec version when they ship one.”

Adoption — the 97M-downloads inflection

The 97M-monthly-SDK-downloads number is from December 2025 Anthropic reporting. The 10,000+ active servers in production and hundreds of distinct AI clients are the same source. The shape of those numbers: MCP went from Anthropic’s connector protocol to the connector protocol used across the AI tooling industry in roughly one year.

The decisive moment for cross-vendor adoption was VS Code + Cursor adding MCP support natively (late 2025), followed by OpenAI’s ChatGPT clients. By mid-2026 virtually every major AI platform or dev tool has some level of MCP support — IDEs (VS Code, Cursor), cloud services and databases, ChatGPT, Claude, Gemini, the lot.

Protocol evolution — Tasks + Server Cards

Two 2026-roadmap additions worth naming:

Tasks primitive

A reliable call-now / fetch-later pattern for long-running operations. Previously: an MCP tool call either returned a result immediately or blocked the agent. Tasks let a server acknowledge the call now and return the result later through a separate fetch — useful for any operation that takes more than a few seconds (model training, large data processing, human-in-the-loop approvals).

Production use has surfaced gaps in the lifecycle semantics: retry semantics (idempotency) and expiry policies (how long results are retained after completion). Both are open issues in the 2026 roadmap.

MCP Server Cards

A .well-known URL convention for exposing structured server metadata without requiring a connection. Equivalent in spirit to robots.txt or humans.txt: browsers, crawlers, and registries can discover what a server does, what tools it exposes, what auth it requires — all from a static URL. This unblocks marketplaces / directories of MCP servers without forcing them to actually call every server.

What this means for Claude Code use

Three practical implications:

  1. Skills that bundle MCP servers are increasingly common. See the skills + marketplace entry — the plugin format’s optional .mcp.json is for exactly this. A plugin shipping a database-query MCP server alongside the skill that uses it is the canonical 2026 shape.
  2. Cross-vendor portability of context-access logic. A well-written MCP server works in Claude Code, Cursor, ChatGPT Desktop, Cline, and whatever comes next. Investment in building good MCP servers compounds across tools the same way skills do.
  3. MCP Server Cards = discoverability. Once cards are widely deployed, finding an MCP server for “thing I want to connect” becomes a search-and-install flow rather than a build-it- yourself flow. This is the 2026 version of “is there a library for that?”

The “USB-C for AI” framing

The most cited 2026 framing of MCP — used in writeups from essamamdani.com and others — is “USB-C for AI-native applications.” The analogy is load-bearing: USB-C is a physical connector that standardises a previously fragmented ecosystem; MCP is a software connector doing the same for agent-to-tool wiring.

The analogy goes deeper than aesthetics. USB-C succeeded because it was vendor-neutral by design (USB-IF governance), backed by the major manufacturers, and good enough technically that nobody could justify maintaining their own incompatible alternative. MCP’s 2025–2026 path has hit the same three checkpoints: vendor-neutral governance (Linux Foundation), backed by the major AI companies (Anthropic, OpenAI, Google, Microsoft, AWS), and good enough technically that the alternatives (function-calling-with-custom-schemas, vendor- specific connector frameworks) have stopped seeing serious investment.

Try this

  1. Read the official MCP roadmap page. Priority-area organisation, working-group structure, the “On the Horizon” section showing what’s getting real community traction.
  2. Look at one MCP server’s source. The reference servers in modelcontextprotocol/servers GitHub repo are small enough to read end-to-end. Pick a domain you care about (filesystem, postgres, git) and walk through how the server declares its tools / resources / prompts.
  3. Install one MCP server in your Claude Code config. The filesystem server is the safest place to start — exposes read-only file access to the model in a controlled way. Watch how its tools show up in CC’s available-tools list.

Sources