Until now, an MCP client could read your product but not change it. That boundary is gone. Any AI assistant you connect — Claude, Cursor, ChatGPT, or any other MCP client — can now run product work end to end.
What it can write
Over a single endpoint, scoped to your org:
- Create and update features, objectives, and key results.
- Plan and adjust sprints.
- Draft and edit pages.
- Cut releases.
- Set up experiments.
- Log decisions.
Reads still work the way they did — this adds the other half, so a conversation can go from "what's the state of X?" to actually moving it.
Parity, no gate
The hosted server and the stdio package now carry the same 71 tools, so it behaves identically whether you connect over HTTP or run it locally. And none of it is plan-gated — write access isn't a tier, it's just how the server works.