MCP overview
The Flax Model Context Protocol server — expose your semantic layer to agents, with all access enforced by the caller's scopes and RLS
The Flax MCP server exposes the platform as Model Context Protocol tools, so an agent such as Claude can discover your semantic layers, validate queries, and build charts and dashboards. This page is for developers connecting an MCP client to Flax.
#What it is
MCP is a thin wrapper over the REST API: each tool maps 1:1 to an endpoint. The agent follows a natural workflow — discover semantic layers and models, validate a query without saving, then create charts, sheets, and dashboards. See the full list in MCP tools.
#Security model
The MCP server adds no privilege of its own. Every call is authorized by:
- the caller's token scopes (
readto explore,writeto create or update), and - the service account's or user's team membership and row-level security,
all enforced server-side by the API. An agent can never see or change anything its principal could not reach directly.
#Two ways to connect
| Mode | Transport | Auth | Use it for |
|---|---|---|---|
| Local | stdio, via the flax-mcp binary |
a service-account API token | your own machine, CLIs like Claude Code |
| Hosted | HTTP, POST /mcp |
OAuth | shared/remote clients with interactive login |
Both expose the identical tool set.