In November 2024, Anthropic released the Model Context Protocol with 100,000 monthly downloads. By March 2026, it had reached 97 million — a 970x increase. Over 10,000 public servers are registered. The protocol crossed the chasm from developer experiment to production infrastructure. But the ecosystem is still maturing, and significant gaps remain.
What is real today
PKM is the #1 use case
Personal knowledge management is the most active segment of the MCP ecosystem. Obsidian has 64+ community servers. Notion has 78+ including an official hosted server. Readwise launched and maintains its own official server. The density reflects a fundamental truth: personal knowledge bases are exactly the kind of rich, structured context that makes AI agents dramatically more useful.
Client support is table stakes
Claude Desktop, Cursor, Windsurf, OpenCode, Hermes, Openclaw, Pi, and Zed all ship first-party or early MCP client support. ChatGPT and Gemini have not shipped first-party MCP support as of mid-2026, though adapters exist. The strongest support remains in coding tools — which makes sense, since developers are the earliest adopters.
HTTP transport is production-ready
While stdio dominates local tools, SSE and HTTP-based transports are no longer experimental. Several cloud providers host managed MCP endpoints. Glassy's stateless HTTP transport — where every JSON-RPC message is a separate POST — has been in production since early 2026 and handles thousands of queries per day without persistent connections.
Open models close the tooling gap
Earlier in 2025, open-weight models lagged proprietary models in tool-use reliability. By mid-2026, Llama 4, Qwen 3, and Mistral Large 3 are competitive on structured output and function calling. This expands the set of hosts that can run MCP servers locally — you are not locked into Claude or GPT-4 for agent workflows.
What is still broken
The mobile gap
Most MCP servers require the host desktop application to be running. An Obsidian MCP server needs Obsidian open on a desktop. Knowledge workers who capture primarily on mobile — a majority of professionals' on-the-go workflow — have no MCP bridge available. The protocol's 2026 roadmap identifies streamable HTTP transport as a priority to address this, but it is a future fix for most implementations.
Glassy is ahead here. Our stateless HTTP transport runs on infrastructure, not as a local process. You can query your knowledge base from any MCP client regardless of whether your desktop is online.
Authentication is still per-tool
Current MCP setups require per-tool API keys and JSON configuration files. There is no SSO, no OAuth flow, no standard auth mechanism. Adding a new server means editing a JSON file and pasting a key. The 2026 roadmap promises SSO-integrated authentication flows, but today, setup is manual.
Server quality is inconsistent
With 10,000+ servers, quality varies wildly. Some are production-grade with rate limiting, audit logging, and least-privilege scopes. Others are weekend hacks that run unrestricted shell commands. Sandboxing, input validation, and least-privilege scopes are now expected — but not enforced. Servers that run unrestricted commands are losing trust.
Graph awareness is rare
Most MCP servers expose flat search — keyword or semantic. Graph-aware servers that understand backlinks, frontmatter, and cross-references are the exception. The PKM ecosystem has rich graph structures (Obsidian wikilinks, Notion relations, Roam blocks), but most MCP servers flatten them into text. Cross-tool standards for graph queries do not exist yet.
What the 2026 roadmap promises
The MCP roadmap identifies three improvements that directly affect PKM use cases:
- Streamable HTTP transport — Moving MCP servers from locally-hosted processes to remote-hosted services. This would address the mobile exclusion problem without requiring a desktop application. Glassy already does this.
- SSO-integrated authentication — Replacing per-tool API keys with standard OAuth connections. This would reduce setup from "edit a JSON file and paste a key" to "click authorize."
- Multi-agent orchestration primitives — Letting specialized agents coordinate across tools rather than requiring a single agent to manage every MCP connection simultaneously.
Predictions for late 2026
- MCP becomes the default integration layer for AI-native SaaS products by Q1 2027
- A certification or compliance badge ecosystem emerges for enterprise MCP servers
- Managed MCP marketplaces appear inside major cloud providers
- Agent-to-agent protocols complement MCP, not replace it
- Open-source server count in curated directories exceeds 500 by end of 2026
Where Glassy fits
Glassy's MCP server is already ahead of the roadmap in two areas:
- Streamable HTTP transport — Live since early 2026. No desktop process required. Remote access from any device.
- Per-user rate limiting — Keyed by MCP token, not IP. Returns JSON-RPC formatted 429 errors. Production-grade.
Where we are still catching up:
- SSO — we use dedicated MCP keys, not OAuth. This is on our roadmap.
- Graph-aware queries — we expose flat hybrid search, not graph traversal. Wikilinks and backlinks are preserved in content but not yet queryable as graph structures.
The most useful frame for evaluating any PKM setup in 2026 is not the server count. It is two prior questions: how does information get into the knowledge base, and how consistently does that actually happen? The AI retrieval layer — whether built on MCP or native to the tool — is only as valuable as what it has to work with. A perfectly configured MCP server traversing a sparse, outdated vault returns sparse, outdated answers.
Glassy solves both halves: the capture pipeline (browser extension, voice, RSS, Obsidian sync) gets information in, and the MCP server with hybrid search gets it out. The loop is closed. Ready to try it? Follow our step-by-step guide to connecting Glassy to Claude Desktop, Cursor, Windsurf, OpenCode, and other MCP clients.