Docs menu
Set up and manage a workspace through MCP — with a person always in the loop.
The Model Context Protocol server lets any MCP-native AI tool read and propose changes to your workspace. It's built so a proposal is the only thing an AI can create — never a live change.
There is no "accept" tool. That's intentional.
Every mutation an AI proposes through MCP surfaces as a proposal a person reviews and accepts in the workspace UI — not through MCP itself. This isn't a policy layered on top; it's how the tool set is built.
- AI can read your workspace, run queries, and draft proposals
- AI cannot accept, finalize, or commit a change — that action doesn't exist as an MCP tool
- Proposals never auto-accept, regardless of which AI client is connected
- A person always makes the final call, in the same review flow as any manual change
MCP sees only what the caller is allowed to see.
Each MCP session is scoped to the caller's access rights — the same rights that govern the workspace UI. Tools and resources outside those rights simply aren't exposed to that session.
Access gating is real; full per-tool controls are still in progress.
Session-level access-rights scoping is real and enforced today. A more granular layer — per-tool rate limits, quotas, and detailed audit logging specific to MCP usage — is actively being built out and isn't fully in place yet. We'd rather tell you that directly than round up.
Read, query, and propose.
- List and inspect data connections
- Run and save warm queries
- Download cold data exports
- Tail live telemetry
- Draft and manage proposals for a person to review