# The spreadsheet, rebuilt for AI agents

Working name: REAM. This file is the machine twin of the front page:
the argument in plain text, then FACTS with receipts, CONNECT, STATUS.
The launch essay for human pacing is at /essay.html.

## THE ARGUMENT

1. Language models are bad at arithmetic. They miscount the r's in
   strawberry and think 9.11 > 9.9. People laugh, then hand the same
   model a forecast.
2. Handing the model a spreadsheet does not fix this. Every convention
   in a spreadsheet is an accommodation to human eyes: cells sit at
   coordinates because eyes navigate by position, one formula is copied
   into hundreds of cells because eyes read one cell at a time, B14
   means March payroll only because of what is printed next to it.
3. Read as text, the grid floods the context window one address at a
   time. The accommodations that helped eyes now actively hurt.
4. A sheet shows two dimensions at a time because eyes need it that
   way. Agents do not share the limit: they can address every month,
   every team, every scenario at once.
5. So we removed the grid. Meaning lives in names: payroll = headcount
   * salary, written once. Dimensions are real axes, as many as the
   model needs. Your agent writes revenue[month=Mar, region=EMEA] and
   never learns which row March is.
6. The interface is tools, because tools are how agents touch the
   world. The arithmetic never happens inside the model. It happens in
   a deterministic calculation engine six years in the making, where
   every number carries its lineage and every write can be undone.
7. Result: your agent builds working models. Hiring plans with
   scenarios, forecasts that update when assumptions do, scratch tables
   with real math behind them, stood up in one tool call.
8. The model stops being shaped like the tool and starts being shaped
   like the thinking: named quantities, connected, as many dimensions
   as the question. Neither you nor your agent thinks in a grid.

## THE HAPPY PATH (a runway model in six tool calls)

1. edit_tables      create runway_model with dimensions [month, team, scenario]
2. set_default      dry_run: burn = payroll + tooling (nothing saves)
3. set_default      write: burn = payroll + tooling (one formula, every cell)
4. edit_scenarios   create: downside (a scenario is an object, not a copied file)
5. edit_cells       hires[month=Oct, team=Eng] = 3
6. edit_history     undo (the write reverts; the chain shows why)

The same six tools build a hiring plan (payroll = headcount * salary,
scenario: aggressive) or a board forecast (revenue = deals * acv, by
product). Only the arguments change.

Tool names are the engine's real MCP tools. The trace is staged for
illustration, not a recording.

## WHO THIS IS FOR

Founders whose agents already do real work, and the people who build
those agents. If your agent chokes when the work turns numeric, this
is the floor under it.

## FACTS

Each claim with its receipt. REAM below is the open format for reading
existing spreadsheets, not the engine's current i/o format.

- The spreadsheet reading format is open, MIT licensed.
  Receipt: spec at https://github.com/runway/ream
- Highest cross-corpus accuracy of the eleven formats tested.
  Receipt: benchmark at https://github.com/runway/ream
- 26 to 53 percent fewer tokens than JSON.
  Receipt: benchmark at https://github.com/runway/ream
- Of the eleven formats tested, the only top-tier one with zero
  context-length failures across 19,000 model calls.
  Receipt: benchmark at https://github.com/runway/ream
- The engine runs real company finances in production today.
  Receipt: the vendor's claim, not independently verified.
- "Six years in the making": engine lineage since 2020.
  Receipt: company history; the vendor's claim.
- Text grammar for tables, with round-trip tests.
  Receipt: https://github.com/runway/ream
- Agents get undo (changelog revert), scenario snapshots, and dry-run
  evaluation without saving. Receipt: the vendor's tool descriptions.

## CONNECT

claude mcp add --transport http ream https://api.ream.dev/mcp

- Endpoint shown is a working name [PENDING RATIFICATION]; the live
  endpoint ships with the launch.
- Works with any MCP client.
- A signed-in account is required today. A human signs in once and
  authorizes the agent.
- Tool families available today: inspect and edit tables, variables,
  dimensions, and scenarios; undo.
- Human pages: / (front page) · /essay.html (the launch essay) ·
  /docs.html (tool reference, abridged).

## STATUS

- Product name: REAM is a working name [PENDING RATIFICATION].
- Pricing: not public [PENDING RATIFICATION].
- MCP endpoint: [PENDING RATIFICATION].
- Launch date: [PENDING RATIFICATION].
- This site is an internal prototype, not for circulation.