Originally published at curatedmcp.com/blog/week-2026-36

MCP Ecosystem Week 36: When Official Integrations Become Your Biggest Governance Challenge

The MCP ecosystem is consolidating around official vendor integrations. This week's usage data shows developers gravitating hard toward GitHub Copilot, OpenAI, Figma, and Anthropic's own MCP servers—a pattern that surfaces a critical governance decision for platform teams: how to manage the blast radius when first-party tools become MCP clients.

This Week in MCP

No new servers were added to the CuratedMCP catalog this week. The catalog remains at 73 risk-classified servers, all free to review. The flat week on new submissions isn't a slowdown—it reflects consolidation. Developers are asking for deeper integrations with tools they already own (GitHub, OpenAI, Figma, Anthropic), not more point solutions. For platform teams, this is a clarifying signal: the governance battle isn't about blocking niche tools anymore. It's about controlling access to the integrations your developers will demand.

On the Radar

GitHub Copilot MCP (98,010 views) and GitHub MCP (76,020 views) lead adoption this week. Both grant Claude and other MCP clients direct access to GitHub repositories, issues, pull requests, and workflows. Before allowlisting: verify your authentication model (token scope, expiration, revocation). If one developer's GitHub token is embedded in an MCP server config, a single compromised machine leaks access to all your repos. Enforce token rotation and consider machine-level enforcement through CuratedMCP's per-device policies.

OpenAI MCP (87,008 views) bridges Claude to GPT-4o, DALL-E, Whisper, and Embeddings. The governance question: are you accidentally creating a dual-model spend pattern? Developers might route work to OpenAI endpoints through Claude without visibility. This is where TokenShield's spend ledger becomes essential—you need a live view of which models are running where before you can govern them.

Figma MCP (82,012 views) pulls design files and component tokens directly into AI workflows. The risk: design tokens and component structures can leak proprietary design patterns. Allowlist this only after confirming your Figma workspace permissions are segregated by team, and that token-level audit logs are enabled.

Anthropic Claude MCP (76,006 views) enables nesting Claude calls within Claude. This multiplies token spend and latency, but also creates reasoning chains that are harder to audit. If you allowlist this, require that nested Claude calls are logged separately and that developers document why they're using sub-agents instead of a single, direct call.

Governance Take

Here's what keeps most platform leaders up at night: developers are already using these servers in Cursor and Windsurf outside your allowlist, and you don't know it yet. Shadow MCP usage is real. The GitHub Copilot and OpenAI numbers this week aren't coincidental—they reflect what developers want to do, not what your policy says they can do.

Start a two-week audit. Query your IDE telemetry, GitHub Actions logs, and your Claude API spend. Where are MCP calls happening? Which authentication tokens are in play? Then decide: do you allowlist the official integrations under tight auth controls, or do you ban them and accept that you're now managing workarounds?

One more thing: if you're running TokenShield, this week is the moment to pull spend data by MCP server and model. You'll likely find that OpenAI MCP and Claude MCP usage is adding up faster than expected. Visibility first, then policy.

---

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