Connect over MCP

D2B’s MCP server lives at https://d2b.dev/mcp/ (Streamable HTTP). Auth comes in two forms: the host’s sign-in (OAuth 2.1 — approve in the browser, and the host renews a one-hour token on its own) or a PAT Bearer header. The tools are the same surface as REST / CLI / SDK, each with a typed schema for tool calling (your host’s tool list is always the current one).

The norms come from the server

On connect, the initialize response carries instructions: the operating norms (which tool when, bring files in by reference, derive with transforms, re-read on 409, commit_snapshot at milestones, delete_workbook to clean up …). If your host truncates instructions, the same text is readable as the d2b://guide resource. Your repo only needs the connection and the choice of path (Use D2B from a coding agent).

Workspace provisioning works over MCP too (provision_workspace / delete_workspace; needs an account-wide key with the workspaces:create / workspaces:delete scopes — a key pinned to one workspace cannot create siblings).

Sign-in (OAuth)

Register the URL and nothing else: the host discovers the authorization server from the 401 on https://d2b.dev/mcp/ and opens D2B’s consent page in the browser. The page shows the app’s name, the host it returns to, the permissions it gets (read / write / delete on workbooks plus cloud-files:read) and which account it acts for (your default account or one of your developer accounts). The access token lasts one hour; the host renews it with a refresh token. It appears under Settings > Tokens in the console as oauth:<app name>, so each connection can be revoked on its own.

The app name is self-declared by the host; D2B does not verify it. Deny a consent page you did not expect. For hosts without OAuth support, and for unattended environments such as CI, keep using a PAT Bearer header.

Per-host setup

Claude Code

claude mcp add d2b --transport http https://d2b.dev/mcp/

Registered without a header, the host asks you to sign in on first use. Unattended environments pass a PAT instead:

claude mcp add d2b --transport http https://d2b.dev/mcp/ \
  --header "Authorization: Bearer $D2B_PAT"

Cursor

~/.cursor/mcp.json (or the project’s .cursor/mcp.json). The URL alone makes Cursor prompt for sign-in; add headers to pin a PAT instead:

{
  "mcpServers": {
    "d2b": {
      "url": "https://d2b.dev/mcp/",
      "headers": { "Authorization": "Bearer d2b_pat_..." }
    }
  }
}

VS Code

.vscode/mcp.json. The URL alone makes VS Code prompt for sign-in (add headers for a PAT):

{
  "servers": {
    "d2b": {
      "type": "http",
      "url": "https://d2b.dev/mcp/",
      "headers": { "Authorization": "Bearer ${input:d2b_pat}" }
    }
  },
  "inputs": [{ "id": "d2b_pat", "type": "promptString", "password": true, "description": "D2B PAT" }]
}

Claude Desktop / claude.ai

Settings → Connectors → Add custom connector with https://d2b.dev/mcp/. Keep the default “Sign in now” (the browser opens D2B’s consent page). To connect with a PAT instead, choose “No sign-in” and set the request header Authorization: Bearer d2b_pat_.... On versions that offer neither, go through mcp-remote:

{
  "mcpServers": {
    "d2b": {
      "command": "npx",
      "args": ["-y", "mcp-remote", "https://d2b.dev/mcp/",
               "--header", "Authorization: Bearer d2b_pat_..."]
    }
  }
}

If you use a PAT, make it least-privilege — workbook-scoped, with a rate limit (Quickstart §1).

Tool map

Purpose	Tools
Discover	list_my_data / get_schema(include_json_schema) / profile_table / get_lineage / get_downstream / search
Read	read_table / query_sql / validate_sql
Review	review_table / review_workbook (evidence-backed findings: the source file’s formulas, the table’s invariants, labels vs. definitions; mode="agent" returns a job — follow it with get_job)
Bring data in (reference first)	ingest_url / request_upload → ingest_upload / list_cloud_files → import_cloud_file / track_onedrive_file / ingest_file
Inspect structure / faithful revise	analyze_source / get_parse_spec / update_parse_spec / materialize_source / revise_source
Containers (workbooks, workspaces)	list_my_workspaces / create_workbook / delete_workbook / provision_workspace / delete_workspace
Create & fix	create_table / add_transform / list_transforms / upsert_rows / delete_rows / write_a1 / put_sheet / get_sheet / list_sheets
Schema	add_column / rename_column / retype_column / drop_column / rename_table
Formulas & descriptions	set_formula_column / list_formula_columns / clear_formula_column / set_artifact_description / set_column_description / set_table_style / get_table_style
History	list_snapshots / commit_snapshot / restore_snapshot / diff_snapshots / recompute_stale / list_ops / undo_op
Conflicts, sync, delivery	list_conflicts / resolve_conflict / export_tables / bind_external_sheet / list_sync_bindings / unbind_external_sheet / list_file_links / sync_file_link

Typing

Tool inputs are strictly typed (enums / structured models, so a harness can reject bad values at generation time). The one data-dependent input, upsert_rows.rows, can be constrained at runtime with the per-row JSON Schema from get_schema(name, include_json_schema=true).

Errors use the same problem+json vocabulary as REST (with suggested_fix) — Reading errors.

The transport is stateless streamable HTTP — no conversation session is kept server-side; each request is self-contained (safe for retries and horizontal scaling).