D2B for Agents

D2B is “Spreadsheets for AI Agents”. It takes messy, human-made Excel at the door, hands your agent typed, row-identified, versioned, governed tables, and returns results in a form humans can read (xlsx, or the original file’s own formatting). Not just answers — every number is traceable to where it came from.

Three ways in

Path	Best for	First call
MCP	MCP hosts — Claude Code / Claude Desktop / Cursor / VS Code	`claude mcp add d2b --transport http https://d2b.dev/mcp/ --header "Authorization: Bearer $PAT"`
CLI	Shell-driving agents, CI, humans	`pipx install d2b-sdk && d2b login`
SDK / API	Your own agents and apps (Python / TypeScript / HTTP)	`pip install d2b-sdk` → Quickstart. PyPI / npm / GitHub

All three are the same surface on the same backend. MCP tools, CLI subcommands, SDK methods and REST endpoints map one-to-one — start anywhere and switch freely.

What makes it different

Traceable: derived tables are built with transforms (SQL / Python); {{ arg }} bindings keep lineage. A code interpreter answers but can’t be traced or reproduced.
Revertible: snapshots, named versions, an op log and undo. Built for agents that make mistakes — restore any point in time.
Governed: column-tag × role mask / deny enforced in the data layer. The same policy applies to SQL, export and MCP alike.
Deliverable: xlsx / csv, write-back into the original workbook with only value cells replaced, live per-row formulas, export = branch → re-import = 3-way merge.
Lives in git: pull transforms, sheets, charts and base tables into a repository, review them in a PR, push them back (Your workbook in git).

Where to go next

Quickstart — mint a PAT and make your first call in five minutes
Core concepts — Workbook / Source / Table / Sheet / Version, optimistic locking, idempotency, governance
Pick a path: MCP / Coding agents / CLI / SDKs / Webhooks
Reading errors — problem+json and suggested_fix
API reference — generated from the OpenAPI spec

Agent-facing indexes: llms.txt / llms-full.txt / openapi.json