Skip to content

Introduction

docsxai walks a running web application, follows written flows, and emits screenshot-rich user documentation. You describe a user task once as a flow-file; the docsxai CLI replays it through headless Chromium, captures clean screenshots, places halos and callouts from the flow's annotations, and renders a publishable doc pack. When the UI changes, you re-run. The docs are a build artifact: the flow is the source and a re-run is the refresh.

docsxai runs in two modes:

  • Calibration is AI-assisted and rare. A host agent - Claude Code, Codex, anything that speaks MCP - explores the live app, picks canonical locators, and authors the flow-file. This happens once per flow, and again only when the app drifts out from under it.
  • Execution is deterministic and continuous. docsxai run replays the flow with zero agent involvement and zero LLM calls. Same flow-file, same target state, byte-identical screenshots - enforced by a keystone test against real Chromium.

The engine never calls a model API. Per-commit LLM runs would be untenable and per-commit Playwright runs are standard CI practice, so you pay for inference once, at authoring time. Every refresh then costs about as much as a test run. Architecture covers how the pieces enforce this.

One name everywhere:

NameWhat it names
docsxaiThe product, the GitHub repo kalebteccom/docsxai, this site, the CLI binary, and the bare npm meta-package (pnpm add -g docsxai@next).
DOCSX_*The env-var prefix; the workspace config is .docsxai.json, schema ids are docsxai/*@N.
@docsxai/*The npm scope: @docsxai/engine, @docsxai/viewer, @docsxai/plugin, @docsxai/backend and @docsxai/skill are published, and @docsxai/plugin-sharepoint is published too, experimental since 0.3.0-rc.3. @docsxai/mcp and the other @docsxai/plugin-* packages are repo-only for now.

When you type a command, it is docsxai. When you install the CLI, it is the bare docsxai package (which wraps @docsxai/engine and ships @docsxai/viewer); when you depend on the library, it is @docsxai/engine. All of it names the same engine.

A screen recorder or a record-and-replay generator captures what you did once; the recording rots silently when the UI moves, and refreshing means re-recording by hand. docsxai is the better tool when the documentation has to stay current without a human in the loop:

  • The flow-file is a reviewable text file.
  • The replay is deterministic enough to gate CI on pixel-level drift (docsxai diff --fail-on warn).
  • A halted step names exactly which locator or assertion broke instead of handing you a stale video.
  • The same calibration loop works from any MCP-speaking agent, so authoring is cheap too.

If all you need is a one-off capture of a UI that will never change, a recorder is less setup. If you need living docs for an app that ships weekly, write the flow once and let CI do the rest.

Made by KalebtecGitHubApache-2.0 licensed