Recalc Turns Spreadsheet Cells into Claude Code and Codex Terminals
By Mr.Xu Community Post
Published:
Summary:Recalc is an open-source workspace that turns spreadsheet cells into terminal sessions running Claude Code or Codex. Each cell can act as an isolated worker, while a lead session coordinates the grid and agents exchange context through cell addresses, named ranges, or an MCP server. Sessions can also be moved into a conventional terminal for continued work. The project includes a small FDA Bench that assembles submission dossiers and tests whether a harness around Claude can predict pass/fail o
The Core Idea Behind Recalc
Recalc turns spreadsheet cells into terminal sessions running Claude Code or Codex, allowing independent workers to exchange context through grid positions, named cells, or an MCP server. A lead session can coordinate the entire grid, while individual sessions can be transferred into a conventional terminal when deeper interaction is required.
The author built the tool for evaluation and benchmark development and is releasing it for free. It currently runs locally on macOS and Linux, although Linux support is less tested, and Windows is not yet supported.
How Sessions, Cells, and MCP Work Together
Recalc uses three mechanisms to coordinate work:
- Grid references let one cell consume another cell's output. This is intuitive but can break when the spreadsheet layout changes.
- Named cells bind outputs to stable identifiers, reducing accidental reference failures.
- An MCP service acts as the connective layer through which sessions exchange information and state.
The spreadsheet is therefore not functioning as a traditional formula engine. It acts as an observable agent orchestration surface: the grid shows task state, terminals execute tools, and models handle code changes and document processing. The MCP layer provides much of the glue, but its exact permissions, message protocol, and isolation boundaries require source-level review.
Operationally, Recalc must manage terminal process lifecycles, multiplex output and keyboard input, and preserve session state. Claude Code and Codex sessions can maintain separate context trees, so sharing only final text may not reproduce the complete state of a running session.
The FDA Bench Example
The project includes a small FDA Bench that assembles FDA submission dossiers and tests whether a harness around Claude can predict pass/fail outcomes. This demonstrates two plausible uses: parallel preparation of unstructured regulatory documents and construction of repeatable document-processing pipelines for model evaluation.
It is not yet a rigorous benchmark, however. The material does not provide enough information about sample size, label construction, data splits, prompts, cost, variance across repeated runs, or accuracy. If labels come from publicly available approval outcomes, models may exploit wording or memorization rather than demonstrate robust document reasoning.
A production-grade evaluation should isolate documents by submission batch, blind the labels, add human review, report confidence intervals, and compare cost and quality against human reviewers or deterministic baselines.
Parallelism Benefits and Costs
Recalc's strongest benefit is operational clarity. It gathers terminal sessions, intermediate artifacts, task state, and delegated outputs into one interface. For benchmark development, this turns the workflow itself into an experimental workbench, making failures and dependencies easier to inspect.
The trade-offs are substantial. Multiple resident coding agents multiply model calls, process memory, file-lock contention, and tool latency. If several agents modify the same repository concurrently, stale reads and conflicting edits can corrupt results.
Shared context also creates propagation risks. Passing an entire cell output into another session increases token cost and may spread erroneous conclusions across the graph. Production deployments need explicit input/output contracts, file ownership, and acceptance checks rather than unrestricted access to the whole sheet.
Practical Adoption Guidance
Start with three workload patterns:
- Parallel retrieval: each cell queries a different source, while a summary cell merges the results.
- Isolated code changes: assign separate sessions to different modules or directories and prevent cross-writing.
- Evaluation pipelines: separate deterministic data preparation, model calls, scoring, and aggregation into traceable stages.
Add process or container isolation, scope credentials and network access per session, impose output-size and token budgets, and log prompts, commands, exit codes, artifact hashes, and model costs. Without those controls, Recalc is an appealing multi-agent workspace rather than a reliable execution platform.
Recalc remains an early-stage project, but its design is worth testing: coding agents become persistent workers in a spreadsheet instead of isolated chat sessions, with MCP and terminal state providing a practical collaboration layer.
— END —Source: Hacker News AI Feed (2026-11-08)
Tags: #Recalc #Claude Code #Codex #MCP #Multi-Agent Orchestration
Editorial & Fact-Checking Note: This article is compiled from primary research, official release documentation, and source papers by the ZICQ Newsroom pipeline with automated entity verification and human editorial review. If you notice any technical inaccuracy, please submit a correction via our corrections policy or email our editorial desk directly.
Community Comments