ALL INTELLIGENCE
Hacker News AIT2·Tools

Open Source Spreadsheet Runs Claude Code Terminals in Every Cell

Original title: Show HN: Recalc – Spreadsheet where cells can be Claude Code or Codex terminals
62AI Score

Recalc is an open source, locally run spreadsheet in which every cell can act as its own terminal, wiring directly into coding agents such as Claude Code and Codex so users can watch agents fill cells in real time and steer them by prompting or editing outputs. The project is in early beta, supports macOS and experimental Linux, and has not been tested on Windows.

Full Report

Contents9 sections

recalc

A spreadsheet as your harness of harnesses, where every cell can be used as its own terminal. Watch agents fill cells in real time, open any one in its own terminal, and steer it by prompting it or by editing the cells it fills.

macOS · Linux (experimental) · works with Claude Code, Codex and pi (experimental) · free and open source · runs on your computer

See it work · Get started · Cells in 30 seconds · Examples · The guide

[!WARNING] recalc is in early beta. It's very, very early in development, so expect bugs, rough edges, and changes between versions that break things. Keep copies of anything you can't afford to lose, and please tell us what breaks.

Why

Spreadsheets are still a great way to organize and analyze data. While it would be possible to just let Claude Code or Codex take control of Excel or Google Sheets, this felt like an opportunity to reimagine the spreadsheet for the agentic AI age from the ground up (though most of it should still feel familiar). This is still very experimental and very much a work in progress.

In its current form, recalc gives you and your agents one shared grid. Every cell can hold a number, a formula, a terminal command, or an AI agent. Agents write their plans and results into cells while you watch. Commands in other cells check their work and turn red when something's wrong. When you want something changed, you edit a cell instead of interrupting the chat. The different Claude Code or Codex instances running in recalc understand how recalc works via an MCP server.

What you can do with it

Check an agent's "done". Put a cell that runs your tests next to the cell where the agent says it's finished. The test cell re-runs by itself whenever the code changes. If the agent says done while a test is failing, you see it straight away. You can even stop a Claude Code session from finishing until its row goes green.

Ask a hundred documents the same questions. Put files down the side and questions across the top, then press Run. An agent answers every cell, and each answer links to the exact passage it came from. Click a cell to see the quote highlighted in the document. Mark the answers you've checked. Agents can't overwrite them after that. Export the result to Excel when you're done, or to share it.

Run a few agents at once without losing track. Give each agent its own row and watch them all on one screen. Anything that needs you (a question, an error, a finished answer) lands in one list called Needs You. Or let the Driver, an agent that works the board for you, set things up from a prompt.

Keep your work on your computer. There's no account, no server and no tracking. The only request recalc makes on its own is a once-a-day check for a new version, which you can turn off. A board is just a folder of files. The agents you connect send what they always send to their own providers. For private documents, you can run the agents on a local model.

See it work

A tour of the app

(1) Every kind of cell, an agent column on Haiku, review and Needs You, the Driver, sheets and sharing.

https://github.com/user-attachments/assets/72b60b54-946b-4236-acd5-4850befe0b27

Contract review

(1) Twelve public contracts and six questions: 72 cited answers, a wrong one caught from its own source, flagged, and fixed by re-running with the note.

https://github.com/user-attachments/assets/705e7993-eddb-477b-ae0f-c0442f809a3d

S-corp returns

(4) Five fictional clients' books carried onto the IRS's draft 2026 Form 1120-S: agents classify each account, formulas build and check each return, and an action fills the PDF only once the checks pass. Run it yourself with recalc demo scorp.

https://github.com/user-attachments/assets/f3d12576-151d-4f5b-a0ab-e836d094aeb4

Tying out a 10-K

(0) Sixteen figures pulled from a public annual report, each with the line it came from, and formulas that check they agree, down to a $14.4 million difference the notes explain.

https://github.com/user-attachments/assets/571af978-53cb-4d08-84ad-3d151bb19a89

A model benchmark

(2) Models predicting FDA decisions from dossiers frozen before each decision: add a new model as a contestant, score it against the others, then open one of its answers in Terminal and ask it why.

https://github.com/user-attachments/assets/6208594e-8072-4a30-9dd7-fd4f22c50499

Get started

  1. Install recalc. In a terminal on a Mac (Apple silicon or Intel):
BASH
   curl -fsSL https://raw.githubusercontent.com/sshah03/recalc/main/install.sh | sh

It puts recalc in your Applications folder and a recalc command on your PATH. It's a short script; read it first if you like. Run it again any time to reinstall. On Linux the same line installs the AppImage, but Linux support is experimental.

<details> <summary>Or install by hand</summary>

Download the .dmg for your Mac from Releases and drag recalc to Applications. The beta isn't signed by Apple yet, so macOS refuses to open a downloaded copy until you clear its quarantine flag once:

Text
   xattr -dr com.apple.quarantine /Applications/recalc.app

(Or try to open it, then click Open Anyway in System Settings → Privacy & Security.) recalc → Install Command Line Tool… adds the recalc command.

</details>
  1. Make sure you have an agent. recalc runs the agents you already use. Install Claude Code, Codex or pi (experimental) and sign in once. Agent → Connect an Agent shows what recalc found.
  2. Try an example. File → New Board… → Examples. "Legal due diligence" is a good first one: select the empty answer cells and choose Cell → Run Selection (⇧⌘↩) to watch agents fill them in.

Your agent can also work on the board from its own chat. To connect a Claude Code you run yourself, run this in your project's folder:

Text
claude mcp add recalc -- recalc mcp

(Without the recalc command, Agent → Connect an Agent shows the same line with the app's full path.) After that, just work as usual: the agent knows the board is there and uses it.

(注:更多技术实现细节与完整 API 文档请查阅原项目 README)


作者发帖说明(HN):

I built this because it was useful for me when building evals/benchmarks, and I decided to put it up for free for others. The goal was quite simple, make it easy to use a lead claude code / codex session in a panel on the side, while allowing every cell to also be a terminal session running claude/codex, and make it possible to continue a session in an actual terminal window if I needed to. They can all talk to each other using cells on the grid, named cells, or the MCP server that effectively glues the app together. I also found it useful for managing context within related but disparate sessions.

My real benchmark runs with recalc are internal to my startup, so to demonstrate some of the functionality I built a small FDA Bench with the tool (obtain and put together dossier's of FDA submissions, and see if a harnessed Claude could have accurately predicted pass/fail): https://github.com/sshah03/recalc#a-model-benchmark

It is still very much a work in progress, but given the versatility of spreadsheet apps, wanted to see how others see themselves using it and what doesn't exist that would be useful to them. Currently runs locally on macs, linux as well but that's less tested, and Windows hasn't been tested at all so that'll come in the future.