More on Sanity MCP:
- Announcement post: sanity.io/blog/sanity-...
- Docs article: mcp.sanity.io
- Agent Toolkit repo: github.com/sanity-io/ag...

6) MCP ecosystem is still early

Getting OAuth2 with refresh tokens working across the various MCP clients (Cursor, Claude Code, VS Code, Lovable, v0, et al.) took multiple iterations. They interpret the spec slightly different, so expect some debugging if shipping a remote MCP.

5) CLIs that auto-configure MCP is becoming the norm

Sanity CLI now detects your editor(s) and offers to configure MCP on `sanity init`. Neon, Sentry, PostHog are doing similar things.

Shoutout to @andrelandgraf.com and Neon team for the inspiration!

4) MCP tool naming matters more than you'd think

Single-purpose tools with descriptive names is preferred by agents. We e.g. improved performance by splitting document creation into `create_document_from_markdown` and `create_document_from_json` rather than one generic tool.

3) Models vary wildly at MCP tool calling

Claude Opus 4.5 is by far the best at this. The gap between models is bigger than I expected – other frontier models struggle with deciding whether to use the MCP at all, or picking the right tools when chaining multiple steps.

2) Agent rules distribution via MCP actually works

We added `list_sanity_rules` and `get_sanity_rules` tools that pull latest rules from our `sanity-io/agent-toolkit` repo on demand. No more manual updating *.mdc files. Agents pick them up just as well as local agent rules.

1) Vibecoding with Sanity feels really nice now

When everything happens through an agent chat – creating project, defining schema and managing content (incl. generating sample content for front-end) – it just flows. No context switching between CLI, Studio, and IDE.

🚀🚀🚀

JD: www.sanity.io/careers/staf...

Source code: github.com/joneidejohns...