Originally published at curatedmcp.com/blog/week-2026-41

MCP Ecosystem Week 41: When Developer Choice Outpaces Your Allowlist

The MCP ecosystem continues its steady climb toward ubiquity in AI-assisted development. This week brought no new catalog entries, but the view counts tell a story platform teams need to hear: developers are gravitating toward integrations that connect their existing tool chains—GitHub, OpenAI, Figma, Anthropic—directly into their coding agents. The question isn't whether your team will use these servers. It's whether you've decided to govern them or discovered them in an audit six months from now.

This Week in MCP

No new servers were added to the CuratedMCP catalog this week. That pause is worth noting. It suggests we're entering a consolidation phase, where platform teams are working through the governance implications of the 80 risk-classified servers already available rather than chasing new additions.

Use this breathing room. If you haven't audited which MCP servers your developers are actually running across Claude Code, Cursor, Windsurf, and GitHub Copilot, now is the time. Shadow usage—servers spinning up without your knowledge—is the governance blind spot we see most often in platform teams shipping AI coding tools at scale.

On the Radar

The five most-viewed servers this week reveal clear patterns in how developers want to extend their AI agents:

GitHub Copilot MCP (98k views) and GitHub MCP (76k views) dominate the list. The first wraps Copilot's own intelligence; the second manages repositories, issues, and workflows directly. Governance consideration: these servers integrate with your source control and CI/CD. Ensure your SSO and RBAC policies enforce per-repository or per-org limits. Token logging matters here—agent-driven GitHub operations can generate audit trails you'll need.

OpenAI MCP (87k views) gives agents direct access to GPT-4o, DALL-E, Whisper, and embeddings. Before allowlisting, clarify your org's multi-model policy. Are developers expected to route through Claude, or can they call competing LLMs? Token spend visibility becomes critical when agents can invoke multiple model providers.

Figma MCP (82k views) opens design files and component tokens to agents. Supply-chain risk consideration: your design system and brand assets flow through this connection. Confirm Figma's auth model (OAuth 2.0, service tokens) aligns with your identity provider. Audit who can access what.

Anthropic Claude MCP (76k views) nests Claude within Claude—useful for specialized reasoning tasks, but introduces recursive LLM calls and compounding token cost. If you're not tracking sub-agent behavior, this server will be a surprise on your monthly bill.

Governance Take

Here's what we're hearing from platform teams in the field: allowlist drift. A security team approves MCP servers for Cursor in Q4. By Q1, developers are spinning up the same servers in Claude Code, and the approval matrix hasn't kept pace. Different IDEs, different enforcement points, no single source of truth.

Add to that the TokenShield gap: most organizations have spend visibility for their primary model provider (usually Anthropic), but the moment agents start invoking GitHub APIs, Figma endpoints, and sub-agent LLM calls, the cost picture fragments. You're paying for tokens. You're also paying for API calls downstream that don't show up in your Claude bill.

The play: build your allowlist once, enforce it everywhere. Map your MCP servers to teams, risk classifications, and cost centers. Log every server invocation—not just token count, but which user, which IDE, which repository access, which external API was called. TokenShield gives you spend visibility and measured, opt-in optimization across your deployed servers. But visibility only matters if your governance layer can act on it.

Govern MCP usage across your team with CuratedMCP — or scan your own stack free at https://www.curatedmcp.com/auditor.