GitHub radar
A spreadsheet that lets you watch and verify your AI agent's work
Recalc is a spreadsheet where every cell can run as its own terminal or AI agent, so you watch agents fill it in live and steer them by editing cells instead of just trusting the chat.
Recalc is a new kind of spreadsheet built for AI agents like Claude Code or Codex: instead of chatting with an agent, you see its plans, answers, and results right in the cells.
When it helps
Useful if you already use an AI agent for code or document work and want to see and test what it does instead of taking its word for it. Skip it if a normal chat with your agent is enough, or you're on Windows.
Pros
- See the agent's work as it happens, not just the final answer
- A test cell turns red if the agent said 'done' too soon
- Free app, your data and board stay on your computer
- Document answers link to the exact line they came from
Cons
- Early beta: expect bugs, keep backups of anything important
- Mac only for now; Windows isn't supported, Linux is experimental
- Needs paid accounts with each AI provider you want to connect
How to set it up — step by step
- 1Open a terminal on a Mac (Apple silicon or Intel) and copy the install command from the Get Started section at github.com/sshah03/recalc.
- 2Run it — recalc lands in Applications and a recalc command becomes available in your terminal.
- 3Make sure you have an agent installed and signed in: Claude Code, Codex, or pi.
- 4Connect your own Claude Code by running claude mcp add recalc -- recalc mcp in your project's folder.
- 5Send your agent the text from the block below so it installs recalc and shows you a first working example.
- 6Check the result: open the board that comes out and watch the agent fill in its cells.
Text for your agent
Copy this and send it to your agent — Claude Code, Codex, any of them:
Install recalc from https://github.com/sshah03/recalc — follow the README's Get Started steps for Mac, connect your Claude Code with `claude mcp add recalc -- recalc mcp`, and show me the resulting board with a first example of an agent at work.
Runs on a regular laptop — needs a Mac (Apple silicon or Intel), no GPU required. Linux support is experimental, no Windows yet.
Open on GitHub▌ More finds