title: "Agents" sidebar_position: 1

Agents

Coding agents and assistants work on .012 workbooks through the same operations people use: every change is typed into cells as the screen types it, checked by the same rules, made as one change and shown by 012 diff like anyone's. Nothing agents do runs a program or reaches the network unless they were started with the flags that allow it, as for scripts.

Page	For
This page	Agents with a shell: the commands, their JSON, and the Claude Code skill
MCP server	`012 mcp`, for any MCP host: tools, resources, prompts, views in the chat, and adding it to Claude Code, Codex and Claude Desktop with `012 agent --install-mcp`
Live mode	The agent working in your running 012 with you: its pointer in the grid, its changes as suggestions you accept or reject, a scope, questions on the context line
JSON output	The schemas of what commands write with `--format json`, and what the MCP tools return

With a shell

Agents that can run commands use the ones scripts use, in this order:

012 describe book.012 --format json                  # what's where
012 get book.012 Sales --format json                 # read what describe pointed at
012 set book.012 D2 '=B2*C2' D3 '=B3*C3' --dry-run   # preview a change as a diff
012 set book.012 D2 '=B2*C2' D3 '=B3*C3'             # make it: one change, saved atomically
012 recalc book.012 --format json                    # check no formula shows an error
012 export book.012 report.html                      # share it as a web page

012 describe lists each sheet with its used range, the row that looks like its header and the column names in it, frozen panes, filters, tables, notebook outputs and linked files sent to it, and charts (numbered, as 012 export --chart names them); pivot tables with their source; notebook tabs with their cells and whether each ran; and the workbook's named ranges:

Sales (shown)  A1:F9, 9 rows x 6 columns, 54 cells, 9 formulas
  header row 1: Date, Region, Units, Price, Total, Share
  frozen 1 row, 0 columns
  table Orders  A1:F9  Date, Region, Units, Price, Total, Share
  chart 1  Units by date  column of A1:C9 at H2
Notebook (notebook)  2 cells
  1 code  files = ls  (ran)
  2 code  $files | where size > 1kb  (not run)
names
  Rate  Sales!B12

A header row is guessed: the first row with contents, when every cell in it is text and the rows below hold numbers, dates, booleans or formulas, or when it's bold. --format json gives the same as the schema.

012 set --dry-run checks the entries as set does (a formula that doesn't parse, a validation rule, a protected range stop it the same way) and prints what would change, as 012 diff would, without saving; --format json gives the changes as records instead. Every command's flags are in Command line.

The Claude Code skill

012 carries a Claude Code skill: a SKILL.md that tells Claude when to reach for 012 and how to use it well (describe first, preview with --dry-run, recalc after changing formulas, never edit the file's JSON by hand).

012 agent --install-skill                          # into ~/.claude/skills/012, for every project
012 agent --install-skill .claude/skills/012       # into this project, to commit with it
012 agent --skill                                  # print it

Installing over a file that isn't this 012's skill asks for --force; installing the same skill again says it's up to date. A newer 012 may carry a newer skill: install it again after upgrading.

Trust

Agents get no more than scripts do. A workbook's notebook cells run only with --notebooks, and then only when the workbook was saved on this computer or --trust says so; JEV functions ask the network only with --jev. Protected ranges refuse changes unless --force is given. Leave these flags to the person asking.