Recalc, a Spreadsheet With Claude Code and Codex in Its Cells, Warns It Is Not a Sandbox

Recalc, a desktop spreadsheet whose cells can be Claude Code, Codex or pi agent terminals, reached 41 GitHub stars six days after its repository was created on Nov. 3. Its author, who posts as sshah03, put it on Hacker News on Nov. 8, where the Show HN drew 28 points and 8 comments.

The numbers are small, and that is the point of reading it now: this is what an agent-in-a-grid interface looks like on day six, before anyone has decided what it should be.

What a cell can hold

According to the project README, a cell is one of five things. A plain value. A formula, starting with =. A shell command, starting with $, that re-runs when the cells it references change. An action, starting with !, that runs only when the user triggers it. Or an agent, written @claude, @codex or @pi, that likewise runs only on command. Cells reference each other with braces, so {A2} pulls in the value of A2.

That makes the sheet a dependency graph with a terminal behind every node. A $ cell behaves like a spreadsheet formula whose function happens to be a shell script, and the README says agents running inside recalc learn how the board works through an MCP server, connected with claude mcp add recalc -- recalc mcp.

Cell prefix	Runs when	Example use
=	Inputs change	Arithmetic on other cells
$	Inputs change	A shell command re-run on new data
!	User triggers it	A one-off action
@claude, @codex, @pi	User triggers it	An agent session

Platform, licence and cost

Recalc is Apache-2.0, built with Rust and Tauri with Node.js for the front end, per the README. macOS is the main target, Linux is experimental and Windows is not supported. The README calls it an early beta, expects bugs and breaking changes, and says Apple signing is still pending, so macOS needs the quarantine flag cleared on the .dmg build. The curl install script is hosted on raw.githubusercontent.com, which is the place to read it before running it.

The app is free. Agent use bills against the user's own Claude or ChatGPT plan, and the dollar figures it shows are estimates at API prices. A budget setting, "Set Agent Budget," caps usage, though the README gives no units or default. The README says there is no account, server or tracking, with one daily update check that can be switched off.

The safety line to read first

The README says: "recalc isn't a sandbox, though. Commands run with your permissions, so run it where you'd run them anyway."

That sentence matters because of what a $ cell does. A shell command that re-runs whenever its inputs change is a standing order, and in a shared board the inputs can come from an agent's output. The README does not say which permission mode an agent cell launches Claude Code in. Claude Code's permissions documentation says that in its default Manual mode, shell commands and file edits require approval, so a cell that runs headless has to either surface those prompts or bypass them. Recalc's README is silent on which, and with 0 open issues on the repository on Nov. 10, nobody has asked in public.

Who it fits

The usage readout is the practical feature: the README says it shows how much of a plan's usage a board took. Someone running several agents against one dataset and wanting to see what the board cost has a reason to try it on a machine with nothing valuable on it.

The next thing to check is the project's first tagged release. It has no version number in the README, and 21 commits in six days is a pace the beta label should be read against.

Related: Claude Code 2.1.318 sent a user's email in a header, and Rivet's agent claims are two more reminders to read the fine print on agent tools.