Spreadsheets already behave like a computational graph - cells referencing other cells.
If you've ever built a financial model, you know the feeling: changing a few assumptions recalculates the entire workbook across many sheets.
MetaCells explores extending this pattern with AI processing. Some cells can call models, process documents or receive/send data through connectors.
The interesting part is that the whole reasoning chain becomes visible directly in the sheet, which reduces the mental overhead compared to many agent frameworks.
Spreadsheets may already be the most widely used programming interface in the world.
For decades people have built complex workflows in Excel and Google Sheets using a simple idea: cells referencing other cells.
One calculation produces a result → another cell builds on it → another transforms it again.
A chain of computation emerges across a sheet - sometimes across many sheets.
If you've ever built a financial model, you probably remember the moment when changing a few assumptions triggered a recalculation across the entire workbook - dozens of sheets updating automatically.
MetaCells explores what happens when the same pattern is extended with AI.
Instead of only formulas, some cells can call AI functions and external tools.
A cell can:
call an LLM
analyze files or images dropped into the sheet
receive or send data via connectors (email, messengers, APIs)
process attachments or documents
generate or explain formulas
return structured outputs that other cells can use
The spreadsheet still behaves like a spreadsheet: cells referencing other cells.
But now some cells perform AI processing, and their outputs become inputs for other cells.
Chains of AI computation can flow across the sheet and across sheets.
In a way it feels similar to a financial model - but instead of numbers propagating through formulas, AI-generated results propagate through the workbook.
External inputs can enter through connectors (emails, messengers, APIs), pass through AI-processing cells, and produce structured outputs or outgoing messages.
Compared to agent frameworks like OpenCLAW, the spreadsheet model provides a more visual representation of what is happening. The processing chain is visible directly in the sheet, which can reduce the mental overhead of building and debugging agent workflows.
Example workflows:
A founder drops customer emails into the sheet → AI cells extract problems customers mention → other cells group them into themes → another sheet generates possible product ideas.
Or competitor analysis → AI cells read websites or PDFs → extract features and pricing → another sheet builds comparison tables → another generates positioning ideas.
Just like traditional spreadsheets, these chains can eventually be collapsed into reusable formulas.
MetaCells is open source and can be cloned and run locally.
GitHub:
(link)
Curious what people here think
Could spreadsheets actually be a practical interface for building AI agent workflows?
What kinds of workflows would you build first?
A bit more context on the idea.
Spreadsheets already behave like a computational graph - cells referencing other cells.
If you've ever built a financial model, you know the feeling: changing a few assumptions recalculates the entire workbook across many sheets.
MetaCells explores extending this pattern with AI processing. Some cells can call models, process documents or receive/send data through connectors.
The interesting part is that the whole reasoning chain becomes visible directly in the sheet, which reduces the mental overhead compared to many agent frameworks.
Spreadsheets may already be the most widely used programming interface in the world.
For decades people have built complex workflows in Excel and Google Sheets using a simple idea: cells referencing other cells.
One calculation produces a result → another cell builds on it → another transforms it again. A chain of computation emerges across a sheet - sometimes across many sheets.
If you've ever built a financial model, you probably remember the moment when changing a few assumptions triggered a recalculation across the entire workbook - dozens of sheets updating automatically.
MetaCells explores what happens when the same pattern is extended with AI.
Instead of only formulas, some cells can call AI functions and external tools.
A cell can:
call an LLM
analyze files or images dropped into the sheet
receive or send data via connectors (email, messengers, APIs)
process attachments or documents
generate or explain formulas
return structured outputs that other cells can use
The spreadsheet still behaves like a spreadsheet: cells referencing other cells.
But now some cells perform AI processing, and their outputs become inputs for other cells. Chains of AI computation can flow across the sheet and across sheets.
In a way it feels similar to a financial model - but instead of numbers propagating through formulas, AI-generated results propagate through the workbook.
External inputs can enter through connectors (emails, messengers, APIs), pass through AI-processing cells, and produce structured outputs or outgoing messages.
Compared to agent frameworks like OpenCLAW, the spreadsheet model provides a more visual representation of what is happening. The processing chain is visible directly in the sheet, which can reduce the mental overhead of building and debugging agent workflows.
Example workflows:
A founder drops customer emails into the sheet → AI cells extract problems customers mention → other cells group them into themes → another sheet generates possible product ideas.
Or competitor analysis → AI cells read websites or PDFs → extract features and pricing → another sheet builds comparison tables → another generates positioning ideas.
Just like traditional spreadsheets, these chains can eventually be collapsed into reusable formulas.
MetaCells is open source and can be cloned and run locally.
GitHub: (link)
Curious what people here think
Could spreadsheets actually be a practical interface for building AI agent workflows? What kinds of workflows would you build first?