Witan

Introduction

Witan Labs builds purpose-built tools that give AI agents mastery over the formats humans actually work in.

Today's agents are remarkable generalists. They can reason, plan, and write code — but the moment they touch a real artifact like a spreadsheet, a slide deck, or a financial model, they fall back on lossy approximations: Python objects standing in for workbooks, plain text standing in for layouts, screenshots standing in for diffs. The output is plausible but rarely production-ready.

Witan replaces those approximations with high-fidelity engines an agent can call directly. One binary per product, structured JSON in and out, designed for agents from the first line of code.

Products

Product What it does
Spreadsheet Read, write, recalculate, lint, and render .xlsx workbooks with full Excel fidelity.

More products are in the pipeline. Each will follow the same shape: a single CLI binary, zero dependencies, purpose-built methods for agent workflows, and structured output suitable for downstream tool calls.

Common principles

Every Witan product is designed around the same handful of ideas:

  • Server-side execution. Agents run their logic against a remote engine that handles the format with full fidelity, not a local approximation.
  • Token-efficient by design. Targeted reads, structured output, and agent-native primitives so a model doesn't burn context on irrelevant data.
  • Built-in verification. Every product ships with a way for the agent to prove the work — access tracking, pixel diffs, semantic lint rules, smart error filtering.
  • Zero dependencies. One binary. No heavyweight installs, no protocol adapters, no environment management.
  • Agent skills included. Each product ships with drop-in markdown skill files for Claude Code, Codex, Deep Agents, and OpenClaw, so your agent inherits the right mental model on day one.

Where to start

Get in touch

Next
Pricing