🔌 AI Integration

There are two ways to point an AI assistant at Narro. Which you want depends on whether the assistant already has a terminal.

If your assistant has shell access

Coding agents — Claude Code, Cursor’s agent, Copilot Workspace — can just use the CLI. They need the documentation, not tools:

Read https://getnarro.com/llms.txt, then write a deck about <topic> to deck.md
and run `npx @getnarro/cli build deck.md` to check it.

Every docs page is available as markdown by appending .md, and https://getnarro.com/llms-full.txt is all of them in one fetch.

Offline or sandboxed? Install the docs:

npm install --save-dev @getnarro/docs

import { getDoc, listDocs, getComponent } from "@getnarro/docs";

listDocs();                    // every page's slug, title, description
getDoc("markdown-mode").body;  // the markdown
getComponent("List").props;    // generated from the TypeScript source

The component data is generated at build time from the types, so it describes the version you installed rather than a hand-maintained copy.

One question, three chunks

@getnarro/docs hands over whole pages, and markdown-mode alone is ~8,000 tokens. When the agent should be asking questions rather than reading manuals, install the same documentation as a docspack:

npm install --save-dev @getnarro/docspack docspack
npx docspack sync
npx docspack ask "which layout values are valid in slide frontmatter"

sync indexes every installed docspack into a local SQLite database. ask answers from it — at most three chunks and 3,000 tokens, ranked by full-text search, with no network call. The index holds every documentation page, one chunk per component, and the catalog of valid names.

Then two lines in AGENTS.md or CLAUDE.md are the whole setup:

Run `docspack ask "<question>"` for documentation on this project's
dependencies. It answers from the installed versions.

npx docspack mcp serves the same index over MCP for assistants that would rather have a declared tool.

If your assistant does not

@getnarro/mcp-server exposes Narro over the Model Context Protocol.

Claude Desktop

~/Library/Application Support/Claude/claude_desktop_config.json on macOS, %APPDATA%\Claude\claude_desktop_config.json on Windows:

{
  "mcpServers": {
    "narro": {
      "command": "npx",
      "args": ["-y", "@getnarro/mcp-server"]
    }
  }
}

Claude Code

claude mcp add narro -- npx -y @getnarro/mcp-server

Cursor, Cline, Continue

Same shape, under whatever key the extension uses for MCP servers — for Cline it is cline.mcpServers in VS Code settings.

Tools

And for @getnarro/video, whose timeline is derived from the narration so every step is an edit to sentences an assistant can make:

Terminal demos

@getnarro/terminal holds a terminal session as text. These five are narro-terminal <command> against a playbook file, and each returns the command’s output.

There is no terminal_record tool, deliberately. Every tool above draws text; record runs the playbook’s commands in a real shell, and the CLI guards it by printing the commands and asking a person. A tool would route around exactly that gate, so an assistant authors, renders and checks a demo — and a person runs it.

Recordings

@getnarro/replay records a web app as real DOM. Recording is a person’s job — it drives a browser through a flow — so there is one tool, and it reads.

Call replay_info before writing <Replay chapter="…"> or <narro-replay start="…">. Those name a chapter as a string, and a name the cast does not have is not an error: seek falls back to the beginning, so the demo plays from the top and reads as a timing bug rather than a typo.

Call replay_review before publishing one. A cast is text that travels, and an email address or an API token typed into a form during the recording is in that text. The rest of the pass — a click that changed nothing, dead air, a scroll nobody could read, type that will be six pixels tall on a page — is about whether the demo is worth showing.

There is no replay_apply, deliberately. video_apply exists because rewriting narration in a video moves every cue after it, so the patch has to be computed. A recording has no such coupling: a finding asks either for a re-record, which is a person driving a browser, or for an edit to the .notes.md track — an ordinary markdown file an assistant edits directly and bakes in with narro annotate.

These edit files on disk. Point them at a directory you are willing to have written to.

Resources

The server also exposes the documentation as MCP resources (narro://docs/<slug>), which clients can cache and read without spending a tool call.

Getting good results

The AI prompts page covers this in detail. The short version:

1. Give it the dialect. The most common failure is a model writing a markdown syntax Narro does not speak.
2. Make it build. narro build deck.md is the pass/fail signal; without it nothing catches a bad prop.
3. Say which package. Presentation, Slide, SlideContent, Notes are in @getnarro/core; Heading, Text, List, and the rest are in @getnarro/shared-ui.