MCP reference
Connect Claude, Cursor, or any MCP-compatible AI agent to certified data through the Möbius MCP server.
Möbius exposes certified data as MCP (Model Context Protocol) tools, so an AI agent can query governed data without SQL, warehouse credentials, or a custom integration. There are two distinct ways to connect an agent, with different credential models and different tool catalogs — pick the one that matches how the agent should act. The Workspace API and Workspace MCP act inside your workspace as you, with your own role (management credential); the Data API and Data MCP serve certified data out through the gate on a scoped data credential. See the custody modes page's glossary of what the product itself prints for how these public names map to identifiers you'll see in the CLI and audit trail.
Two connection modes
Service token (below) gives an agent a fixed, workspace-scoped set of 11 data tools — the same posture regardless of who's asking. AI tool session gives an agent your own login identity and role, with a tool catalog gated by a per-environment Limited/Full dial. Use a service token for shared automation that should outlive any one person; use a session for an agent acting on your own behalf.
Data MCP — service token
{
"mcpServers": {
"mobius": {
"url": "{{mcpBaseUrl}}/v1/mcp",
"headers": {
"Authorization": "Bearer <your-integration-token>"
}
}
}
}Generate the token from Settings → Integrations — the same service token works across REST,
GraphQL, and MCP (see REST API reference). The token is
workspace-scoped, so the server only exposes the tables and data groups your token holds
tables:<name>:read or tables:<name>:write on.
Every tool call enforces the same governance gates as REST and GraphQL: only CERTIFIED tables
are queryable, PII fields are masked per your semantic contract, every call is audit-logged with
the calling token's identity, and an in-place table's reads proxy transparently through its
warehouse connector — the agent never sees the difference.
Tools (service-token door)
Eleven tools: six reads and five writes. Tool names match the server's registered tool names exactly.
| Tool | What it does |
|---|---|
list_certified_tables | List certified tables the token has access to |
get_table_schema | Column definitions, types, and semantic contract metadata for a table |
query_table | Read rows from a certified table (an in-place table proxies through your warehouse; a held table reads from Möbius) |
list_data_groups | List the data groups the token can read, with each group's status (all-or-nothing on member-table read scope, same rule as GraphQL). A data group is approved, not certified: its member tables are what is certified |
query_data_group_view | Query a decision view of a data group — pre-joined, governed SQL over multiple certified tables. If the group's views can't be served, the call returns NOT_FOUND ("Data group not found or not certified") |
query_predictions | Read the governed ML predictions ledger for a table. inputSnapshot is withheld unless requested, and PII-masked when it is |
insert_rows | Insert one or more rows into a certified table. Requires tables:<name>:write; gate-evaluated before write; rejected rows are recorded, never silently dropped |
update_row | Partial update (PATCH) to a single row. The merged row is gate-evaluated before the update |
replace_row | Full replace (PUT semantics) of a single row — omitted columns reset to NULL. Replace-existing-only: an unknown row id returns NOT_FOUND, it never creates |
upsert_rows | Batch upsert by key — insert on a new key value, full-replace update on an existing one |
delete_row | Soft-delete a single row (deleted_at is set; the row is never hard-deleted). Re-deleting an already-deleted row returns NOT_FOUND |
Write tools are refused for in-place tables (data in your own warehouse — Möbius doesn't write into it on your behalf) and for vector-shaped tables.
This door is deliberately data-only — it exposes no tool that certifies, approves, or otherwise mutates governance state, and no tool that returns full certification evidence (only the verdict, via each response's governance block). Governance actions run through the CLI, the Workspace API, or your Möbius session in the dashboard, never through this door.
Governance block on every response
Every tool response that reads a table carries a governance block — the same verdict shape
described in the GraphQL reference:
certification status, gate mode, freshness, and which fields were PII-masked. query_table
responses additionally carry a per-row _governance_flag (underscore-prefixed so it can't collide
with a data column of the same name) marking any row that was admitted under an advisory gate with
a failed predicate — disclosed, never filtered.
Workspace MCP — as yourself
From Settings → Integrations → MCP, choose Connect your AI tool and sign in. This starts a Möbius session for your AI client (Claude, Cursor, a script) that acts as you — your live role across data, definition, and control, pinned to one environment chosen at sign-in. Manage or revoke it any time from Settings → Integrations → Live sessions.
Because this door can act with your permissions, its tool catalog is much broader than the 11 service-token tools — it can reach schema and dictionary edits, contract authoring, and other definition-plane work your role allows. Whatever it returns leaves to a third-party AI provider Möbius doesn't control, so your workspace decides how much of that catalog is allowed to leave at all:
- Limited (the product default) — the conservative, external-consumer posture: certified data reads plus schema/dictionary reads, PII-masked. Nothing else. No raw or uncertified rows, no governance evidence, no mutations.
- Full — the whole catalog your role permits, including uncertified data, unmasked content, and writes.
This is set per environment on Workspace → Environments, and the value shown at Tools → MCP Data Control is the resolved posture for the environment your session is pinned to. A workspace opts up to Full deliberately — Limited is always the starting point.
Health check
GET {{mcpBaseUrl}}/v1/mcp/healthUnauthenticated liveness check — confirms the server is reachable and reports whether it's running the current schema version. No workspace data is exposed.
Errors
Tool-execution failures (a forbidden scope, a not-found row, a license restriction) come back
inside the JSON-RPC result with isError: true, rather than as a transport-level failure — so
your MCP client or model sees the error text directly. A missing or invalid token is refused with
an HTTP 401 before any tool runs.
Protocol
protocolVersion: "2024-11-05" (negotiated; echoes what your client requests)Supports initialize, ping, tools/list, and tools/call over both the stdio and HTTP JSON-RPC
transports. The server advertises the tools capability only — it exposes no MCP resources or
prompts.