Run Claude Code and Codex sessions as spreadsheet cells for evals
Author built a spreadsheet where each cell is a Claude Code or Codex terminal, used to run benchmarks such as a demo FDA submission pass/fail prediction bench. ·
Done withClaude CodeCodex
- What they did
- recalc is a desktop spreadsheet where a cell can hold a Claude Code or Codex agent, written as @claude(...) or @codex(...) with references to other cells. Agent cells run only when the user runs them, and an MCP server tells the agents how the board works. In the FDA demo, models predict FDA decisions from dossiers frozen before each decision. A new model is added as a contestant and scored against the others. The user can open one answer in a terminal and ask the model why.
- How it went
- The source describes the FDA benchmark only as a 2:37 demo video. It states no accuracy scores or run times. The project is an early beta and may have bugs.
- Worth knowing
- Agents run on your own Claude or ChatGPT plan. recalc shows each run's cost at API prices, and Set Agent Budget caps usage. It runs macOS-first, with Linux experimental.
Try it yourself with Claude Code
Build me a simple local web app that shows a grid where each cell runs its own terminal session of a coding agent on one task from my eval set at [path to folder or file of test cases]. Each cell should show live output and a final pass/fail result compared against the expected answers in [file with expected answers]. When done, I should be able to run the grid with one command and see a summary of pass rates, and you should show me the plan before you install anything or run commands that delete files.
Want help getting this working? Join the Vibe Coding Academy ↗ Run by Alex Finn, who also runs this site.
Discussion on HN · Nov 8, 2026
Related use cases
Turn tickets into merged pull requests
Pulls tasks from GitHub Issues or Linear, then runs Claude Code or Codex in isolated Kubernetes pods to carry each one through to a PR.
Build a racing game from a single prompt
Codex made a racing game with multiple racers, eight maps and items from one prompt, acting as designer, developer and QA tester by playing it.
Generate custom interactive HTML lessons for students
A teacher gives Claude Code a passage, level, vocab list and goals, and gets back a full lesson file with flip cards and comprehension questions, replacing hours of weekly worksheet work.
Build a C compiler with a team of parallel agents
Anthropic's engineering write-up on tasking Opus 4.6 agent teams with building a C compiler.
Migrate 180 microservices to a new framework version
Brex built Cursor-based self-correcting agents that let one engineer migrate 180 microservices to a newer Micronaut version in about a month.
Build a game guide chatbot and browser game with Replit Agent
A user built a chat bot app guiding a video game, then a browser game, with Replit; the agent helped but made many extra unrequested changes.
Get 5 like this every morning
The best things people got an AI agent to do, each with the prompt to try it.