Meet Blume: An Open-Source, Zero-Config Documentation Framework That Exports AI-Ready Documents From a Markdown Folder

Hayden Bleasel, a professional engineer from OpenAI, has been released Blumeopen source documentation framework. Blume was ported to npm as version 1.0.3 on the same day. It's as simple as Drop Markdown into a folder and send the document to the site. No application boilerplate is written or stored afterwards. This project is MIT licensed and open source.

What is Blume?

Blume is a command line tool paired with a documentation library. It reads a folder of Markdown or MDX files. From that folder, it generates a production-level documentation site. That output sends navigation, search, headers, and Open Graph graphics. Configuration is always optional and is added one file at a time. The code is a TypeScript monorepo; The published package lives on packages/blume. Blume's documents, below apps/docsit was created by Blume himself. Requires Node.js 22.12 or newer. Works with Bun, pnpm, npm, or thread.

How does Blume work?

Under the surface, Blume produces and drives the hidden Astro project. First, the CLI loads blume.config.ts and scans your content into a graph. Next, it lists the Astro project as a .blume/ directory. Astro then translates each page with a single host route. That route imports Blume's exported components, generated data, and output. In each run, .blume/ regenerates, and only changed files are overwritten. As a result, hot reloading is always faster during programming. The main theme does not support the client JavaScript framework. As a result, pages score well in Core Web Vitals by default. When you need full control, blume eject develops the runtime into a stand-alone Astro operating system. That released project is still dependent on blume the parcel.

Getting started

Setup takes one command, so onboarding is short.

After that, blume dev starts a hot reload server. Currently, blume build writes static HTML and local search index to dist/. The configuration file is the actual TypeScript, validated by the schema.

// blume.config.ts
import { defineConfig } from "blume";

export default defineConfig({
  content: {
    sources: [
      { type: "filesystem", root: "docs" },
      { type: "notion", database: process.env.NOTION_DB },
    ],
  },
});

Because the configuration is typed, programmers catch errors before the configuration is executed. The CLI covers a full lifecycle beyond those basics:

AI-Ready by Design

Besides human readers, Blume targets agents as well. Each page returns raw Markdown when compiled .md in its URL. One flag comes out llms.txt again llms-full.txt agents. Each page can be copied as Markdown or opened in ChatGPT, Claude, or v0. The in-page Ask AI assistant answers the reader's questions directly. It runs on the AI ​​SDK through Vercel AI Gateway, OpenRouter, Inkeep, or any OpenAI compatible endpoint. Blume may also host a Model Context Protocol (MCP) server. With it, Claude Code, Cursor, and VS Code read documents directly.

claude mcp add --transport http your-docs 

That server presents four read-only tools: search_docs, get_page, list_pagesagain get_navigation.

Use Cases with examples

Those skills indicate tangible jobs. For API product, specify OpenAPI or AsyncAPI spec. Blume then provides a reference to interoperability with schemas, auth, and the application playground with Scalar. To find the library, point Blume to your GitHub release. Each release feeds into a change log timeline generated by the RSS feed. For a global audience, add translated files for each language. Blume supports 36 regions, location-aware routing, and right-to-left architectures. For mixed content, combine local files with MDX, Notion, or Sanity. All sources offer the same components.

Comparison with Blume

In context, here Blume argues against three common ways of writing texts. Features change quickly, so make sure the information is current before you use.

Strengths and Weaknesses

Power

• Zero-config startup: the Markdown folder becomes a full site.
• The first static output ships with no core JS client, which supports Core Web Vitals.
• Built-in AI areas: llms.txtMarkdown for each page, MCP server, and Ask AI.
• Type-safe config, so programmers flag errors before the build starts.
• Release the path to Astro's standalone app to reduce long-term lock-in.

Weakness

• Version 1.0.3 is new, so the ecosystem is still young.
• Requires Node.js 22.12 or newer, some locations don't have it.
• Request-time features such as Ask AI and MCP require a server adapter.
• A self-contained model means you own the statistics and strings of the assistant.
• It has fewer third-party integrations than older managed platforms today.

Key Takeaways

• Blume turns a Markdown folder into a zero-sum production documentation site.
• It drives a hidden Astro and Vite project and can export to standalone Astro.
• AI includes a built-in ship: llms.txtMarkdown for each page, and MCP server.
• It is MIT licensed, requires Node.js 22.12+, and reached npm v1.0.3 on launch day.
• Early adopters include Quiver, from Mintlify, and Neon's add-mcp documentation.

Check it out GitHub Repo. Also, feel free to follow us Twitter and don't forget to join our 150k+ML SubReddit and Subscribe to Our newspaper. Wait! are you on telegram? now you can join us on telegram too.

Need to work with us on developing your GitHub Repo OR Hug Face Page OR Product Release OR Webinar etc.? Connect with us