AI documentation

Ask your coding agent (Grok, Claude Code, Codex, Cursor, OpenCode) to write docs. Docsmith gives it tools. Docsmith does not call an LLM and does not need its own API key.

Quick start

docsmith install:ai

Then in the agent:

Write usage docs for this project. Put a quick start on the landing page.

Ask for screenshots or a demo video only if you want them:

Add cropped screenshots and a short demo video of the UI. If you cannot boot a demo, ask me for a URL.

install:ai detects agents you have installed and writes:

  • .ai/skills/docsmith-docs/SKILL.md (also copied into each agent's skills dir)
  • .mcp.json for Claude Code, Cursor, Gemini CLI, Junie, Boost
  • .grok/config.toml and .grok/skills/ when Grok is installed
  • .codex/config.toml when Codex is installed
  • opencode.json when OpenCode is installed
  • .agents/mcp_config.json when Antigravity is installed
docsmith install:ai --agents=grok,claude --source=./app --docs-source=./docs-source --force
docsmith install:ai --no-mcp        # skills only
docsmith install:ai --no-skills     # MCP config only

What the agent should produce

A small set of usage pages, not one page per source file.

The landing page opens with a quick start and the most common use case. The agent does not take screenshots or record videos unless you asked. When you do, the widget is cropped with a little space around it, and login is not in the recording unless the page is about login.

The agent uses read_source, write_markdown, and build_site. It uses capture_media only when you asked for images or videos. See MCP Server.

Build the site

docsmith build --source=./docs-source --output=./docs --title="My App"

You can also start the MCP server yourself:

docsmith mcp:serve --transport=stdio --source=./my-app --docs-source=./docs-source

Old AI pipeline

Earlier versions called an LLM from Docsmith (agent:run, providers, a reviewer). That is gone. Coding agents do the writing with their own keys. The old code is in git history on feat/AI. See Future Scope.