D2B launched a new infrastructure layer designed to bridge the gap between human-centric spreadsheets and machine-executable data operations. The platform exposes typed, row-identified, and versioned tables specifically optimized for AI agents, moving beyond the ephemeral nature of sandboxed code execution. By integrating directly with Model Context Protocol (MCP), CLI, and SDKs, D2B allows agents to ingest, transform, and deliver data while maintaining strict lineage and undo capabilities.

From Sandbox Fragility to Typed Lineage

Traditional agent workflows often rely on letting LLMs write and execute Python in isolated sandboxes. While this produces an answer, it fails to provide a traceable execution path or reproducible state. D2B addresses this by recording all agent work as structured transforms with full lineage. If an input changes, the system recomputes the dependencies, and every historical version remains restorable. This ensures that agent-generated numbers are not just plausible, but provably derived from specific source data.

Governance and Git-Native Workflows

The platform embeds governance primitives directly into the data layer, including least-privilege PAT scopes, idempotency keys for mutations, and concurrent write detection. A standout feature is the 'Workbook as Code' capability, which allows teams to sync transforms, sheets, and charts to Git repositories. This enables standard pull request reviews for AI-written SQL or Python logic before it impacts production data. The system supports major agent frameworks, including Claude Code, Codex, Cursor, LangGraph, and OpenAI Agents SDK.

Pricing and Integration Model

D2B employs a metered pricing model based on data operations and storage, explicitly decoupling infrastructure costs from LLM inference. Calls from external agents using their own LLMs incur no LLM charges, with pricing set at $6 per million operations and $0.18 per GB-month. The company notes that reads are currently unbilled, while one operation is defined as a single write-side action such as a row write, transform run, or commit. This structure aims to lower the barrier for high-volume, low-latency agent interactions.

Key Takeaways

  • D2B replaces fragile sandboxed execution with typed, versioned tables accessible via MCP.
  • Full data lineage and undo capabilities allow for reproducible agent workflows.
  • Git-native syncing enables human review of AI-generated SQL and Python transforms.
  • Pricing is metered by operations and storage, with no LLM charges for external agent calls.
  • The platform integrates with major tools including Claude Code, Cursor, and LangGraph.

The Bottom Line

D2B solves the 'black box' problem in agent data workflows by making every calculation traceable and reversible. For teams relying on AI for financial or operational modeling, this governance layer is the missing link between LLM creativity and enterprise reliability.