Md Business

md-business

[CI] [License: MIT] [Node.js 24 LTS] [pnpm] [Chrome Web Store] [VS Code Marketplace]

⬇ Download the desktop app (Windows / macOS) — installers, no build step. Or add the Chrome extension to read and print Markdown business documents in the browser, or the VS Code extension to validate and preview them where you write code.

English | 日本語

AI-native business documents, powered by Markdown as the source of truth.

md-business is an open-source framework for managing business documents as Markdown files with structured frontmatter, schemas, validation, Git history, and human-friendly viewers.

It is not just a Markdown viewer, a PDF generator, or a template collection.

md-business treats Markdown as the canonical source of truth for invoices, test specifications, design documents, API documents, database design documents, estimates, meeting notes, and other operational documents. AI agents can create, edit, validate, diff, and render these documents directly, while humans can review and consume them through familiar interfaces such as PDF, Google Workspace, Chrome, VS Code, PWA, desktop apps, or GitHub.

Why this exists

Most business documents still live inside tools designed primarily for humans:

• Word processors
• Spreadsheets
• PDFs
• Cloud drives
• Office suites
• SaaS back-office tools
• Internal web forms

These tools are useful for human input and review, but they are not ideal as an AI-operable source of truth.

AI agents can read and summarize these documents, but they often cannot safely and continuously operate them as structured business assets. They struggle with reliable editing, validation, diffing, version control, automation, and handoff across multiple agents or systems.

md-business starts from a different premise:

Business documents should be stored in a format that AI agents can operate directly.

Markdown provides the human-readable body. YAML frontmatter provides machine-readable metadata. JSON Schema provides validation. Git provides history, review, branching, and rollback. Renderers and viewers provide human-facing output.

Core idea

Markdown source of truth        ↓Schema validation        ↓AI agent editing        ↓Git diff / review / approval        ↓PDF / Sheets / Docs / browser / desktop / SaaS integration

The important layer is the Markdown source.

Viewers and integrations are only access points for humans and existing workflows. They exist so that non-technical users can review, approve, print, share, or submit documents without needing to understand Markdown, Git, or AI agents.

What md-business is

md-business is:

• A schema-driven business document system
• A Markdown-first source-of-truth layer
• A document format designed for AI agents
• A Git-friendly alternative to opaque office files
• A bridge between AI workflows and human-facing business documents
• A foundation for local-first and agent-driven business operations

What md-business is not

md-business is not:

• A Notion clone
• A Google Docs clone
• A spreadsheet replacement
• A simple Markdown viewer
• A PDF-only generator
• A closed SaaS workflow tool

Existing office suites and SaaS tools can still be useful as review surfaces, delivery channels, or regulated backends. md-business focuses on the source-of-truth layer that AI agents can actually operate.

Why Markdown

Markdown is a practical base format for AI-native business operations because:

• AI agents can edit it reliably
• Humans can read it without special software
• Git can track every change
• Pull requests can be used for review and approval
• Frontmatter can hold structured business data
• JSON Schema can validate document correctness
• Renderers can produce PDF, HTML, Sheets, Docs, or other outputs
• Local repositories can be operated by desktop tools or autonomous agents

Human role

md-business does not remove humans from responsibility.

It moves humans away from repetitive document manipulation and places them at the review, approval, exception handling, and business decision layers.

AI agents operate the source files. Humans approve the outcomes. Viewers provide the interface.

Example use cases

Test specifications

AI agents generate and update test cases as Markdown. Schemas validate required fields. Git tracks changes. A Google Sheets or browser viewer allows non-engineers to execute and review test results.

Design documents

AI agents maintain basic design documents, API specifications, database design documents, and implementation notes as version-controlled Markdown. Mermaid diagrams, structured frontmatter, and schema validation keep the documents reviewable and automatable.

Invoices and estimates

AI agents update dates, line items, tax fields, customer information, and payment terms in Markdown. Schemas validate required fields. PDF renderers generate human-facing documents. External accounting tools can be treated as integration targets rather than the primary editing surface.

Distribution channels

md-business provides multiple human-facing viewers and automation channels:

These channels are not the core product. They are adapters around the Markdown source of truth.

Current templates

Using a template

Every shipped template uses dummy data — no real company names, registration numbers, bank accounts, or invoice numbers. Copy the file into your own workspace and edit the values in place.

Invoice templates (templates/invoice/):

• standard.md — English baseline, single item at 10%
• standard-ja.md — Japanese baseline, mixed 10% / 8% tax rates, bank account, and 適格請求書発行事業者 registration number (T + 13-digit dummy)
• tax-exempt-ja.md — 免税事業者 (tax-exempt sole proprietor) variant — no registration number, with the 2023–2029 経過措置 footnote
• inbound-eligible.md — Multi-item invoice exercising both the standard and the reduced tax rate
• quote-ja.md — Quotation (種別: 見積書) with a validity date instead of a payment due date
• receipt-ja.md — Receipt (種別: 領収書) with a 但し書き line and an opt-in revenue stamp box

Open any of these in the Chrome extension viewer or render via @md-business/renderer-pdf to produce a print-ready PDF without writing any code.

Database design documents (db-spec/v1 for relational, nosql-db-spec/v1 for document stores) have schemas and open in every app, but no blank template yet — examples/ has a filled one to copy from. Future templates may include meeting notes, contracts, operational checklists, and other business documents.

Examples

Templates give you a blank form. examples/ gives you a filled one.

It holds a single fictional project — an internal equipment lending tool — written out end to end: design document, API specification, database design, test sheet, invoice, and an investigation report with its evidence files. The documents cross-reference each other, so you can see how they connect on a real project.

Company names, registration numbers, bank accounts and employee codes are all invented. Copy the shape, not the values. See examples/README.md.

Screenshots

Screenshots will be added under docs/screenshots/. See docs/screenshots/README.md for capture specs (1280 × 800 px / PNG) and the list of planned shots:

• Chrome extension viewer rendering a qualified invoice
• Chrome extension A4 PDF preview
• Google Workspace Add-on test-spec sheet view
• Google Workspace Add-on "GitHub push" sidebar action

For now, see the published Chrome Web Store listing for the viewer screenshots: https://chromewebstore.google.com/detail/lmdplkkfmgapnhombimeohjliinifgjh.

Forking & customization

md-business is MIT-licensed and explicitly designed to be forked for in-house, consulting, and per-vendor deployments. See docs/fork-guide/ for the recommended fork strategies, upstream sync workflow, per-package customization points, and distribution channel notes.

Long-term vision

md-business aims to make business documents operable by AI agents.

Instead of placing AI on top of opaque SaaS screens, PDFs, spreadsheets, and office files, md-business places structured Markdown at the center of the workflow.

This enables a different kind of organization:

• Documents are version-controlled
• Changes are reviewable
• AI agents can work directly on source files
• Humans approve through familiar views
• Local repositories can become operational workspaces
• Back-office and upstream engineering processes can be automated without locking the source of truth inside a specific SaaS

The goal is not to replace every business tool.

The goal is to make the canonical business document layer open, portable, reviewable, automatable, and AI-native.

Documentation

Where a spec document exists, it includes the JSON Schema link, top-level field table, YAML frontmatter sample, and validation examples.

Desktop app

The desktop app (Tauri 2 + SvelteKit) opens every business-document type above in a purpose-built viewer, edits Markdown in a live split view, and exports A4 PDF.

• Find: type in the file-list filter to search titles and bodies across the folder's Markdown and test sheets, or switch the list to a thumbnail grid.

• Read: zoom the preview (50–300%, Ctrl/Cmd + + / - / 0 or Ctrl + wheel); printing and PDF always come out at 100%.

• Draw: zumen blocks become architecture diagrams, chart blocks draw bar / line / pie charts from a neighbouring TSV file, and data blocks quote a TSV file as a table. The VS Code extension draws the same blocks; the Chrome extension draws zumen only, because it cannot read neighbouring files.

• Stay current: the app checks for a newer release on launch and installs it when you agree. Windows (x64 / ARM64) and macOS builds are published.

• Download (Windows / macOS): https://meta-taro.github.io/md-business/download/

• User guide (English): https://meta-taro.github.io/md-business/manual/

• 操作マニュアル (日本語): https://meta-taro.github.io/md-business/manual/ja/

Built-in MCP server

The desktop app is where the source-of-truth layer becomes operable by an AI agent. It starts a local MCP server on launch, so an AI client can read, validate, search, create, and update the business documents in the folder you have open.

• Connect from the MCP tab: Add settings to the open folder writes .mcp.json for clients that read it, or Copy client settings puts a ready-to-use mcpServers entry on the clipboard.
• The connection details persist across restarts, so the settings you paste once keep working.
• about_md_business answers "what can I do here" with the tools to call, in order, for each task. The document tools are list_schemas, get_schema, read_document, validate_document, create_document, update_document and search_documents, plus read_tsv, append_tsv_row and update_tsv_row for test sheets. Every call is validated against the declared schema and listed in the app.
• Test sheets are the tab-separated format, not Markdown, so the agent fills them in row by row: rows it did not touch stay out of the diff. search_documents lists them alongside the Markdown documents, so the agent can find a sheet without being told its path.
• When the folder is a git repository, git_status, git_diff and git_commit let the agent review what it changed and record it — so a day of edits ends as readable history rather than an untracked pile.
• External JSON / XML (exchange-format invoices, bank statements, accounting exports) is read-only: read_data reads it as a tree and data_to_table quotes a repeating part as a table. read_har reads network captures.
• Logs are never read whole: search_lines and read_lines find and read the lines around a hit, filter_records and aggregate work on JSONL / TSV records, and build_timeline merges several files in time order. Every result is masked for tokens, cookies and email addresses. save_evidence keeps an extract as a file an investigation report can cite, and check_tsv checks test sheets without reading them row by row.
• For a folder declared as a website (md-business.yml), declare_web_mode, web_mode_status, list_site_files, read_site_file and write_site_file let the agent edit the site's own files while the app shows the result live.
• open_in_app puts the document the agent just wrote in front of you, and list_open_documents / close_document let it see and tidy the tabs.
• export_pdf asks the app to open a document and bring up the print dialog, so the agent can hand you a finished A4 page to save — where it goes is still your choice.
• capture_window photographs the app's own window, so the agent can see what it actually renders instead of guessing from the file. Only this window is in the picture — whatever else is on your desk stays off it — and the picture is still yours, so it goes nowhere public unless you say so.
• Reach is limited to the folder currently open in the app, the server binds to loopback only, and requests without the matching token are rejected.
• When the agent writes, the file list and preview follow along on their own — no reload, no manual refresh.

git push is deliberately not exposed as a tool, and the server never receives secrets or API keys. See packages/mcp-server/ for the protocol details.

Release history

The four apps — desktop, Chrome extension, VS Code extension and Workspace Add-on — are versioned independently. CHANGELOG.md lists the current version of each and links to its own history.

Conventions

• Data cell convention — Empty cells are the only canonical representation for unfilled data. Do not fill table cells with — / – / ― / N/A / TBD.
• Snapshot tables — Where fetched numbers live and which columns they carry. Every row records when and where it came from; re-fetched figures are appended, never overwritten.

Quick Start

Use it

No build step, no Node.js, no package manager.

1. Download the desktop app (Windows x64 / ARM64 .msi / .exe, macOS .dmg) and install it. Or add the Chrome extension if you only need to read and print documents in the browser.
2. Open a folder. Any .md file with a schema: frontmatter key opens in the viewer for that document type; .tsv test sheets open in the grid editor.
3. Start from a template — copy one out of templates/, or create a test sheet from a column preset with the ＋ button in the file list.
4. Print to A4 PDF from the preview.

The user guide (日本語) covers the rest.

To let an AI agent edit those documents, point your MCP client at the server the desktop app runs. Open the MCP tab in the right-hand panel: Add settings to the open folder writes .mcp.json there — clients that read it (Claude Code among them) pick it up on their own — or Copy client settings puts the same entry on the clipboard for a client configured elsewhere. Either way it looks like this, with the port and token already filled in:

{  "mcpServers": {    "md-business": {      "type": "http",      "url": "http://127.0.0.1:<PORT>/md-business",      "headers": { "Authorization": "Bearer <TOKEN>" }    }  }}

<PORT> and <TOKEN> are shown in that panel and persist across restarts, so the settings you paste once keep working. The token is an access credential: in a Git repository the app adds .mcp.json to .gitignore when it writes it. The server binds to loopback, rejects requests without the matching token, and reaches no further than the folder you have open.

To run the server standalone instead — without the desktop app — see packages/mcp-server/.

Develop

Requires Node.js 24+ and pnpm (via corepack). npm and yarn are not supported. Building the desktop app additionally needs the Tauri 2 prerequisites (Rust toolchain and the platform WebView).

corepack enablepnpm installpnpm dev

Checks — all of these must pass before a change lands:

pnpm lint       # svelte-check (desktop) / tsc --noEmit (everything else)pnpm typecheck  # same checkers, run as their own taskpnpm test:run   # Vitest, with coverage where it is configuredpnpm build      # every package and app

See CONTRIBUTING.md for details.

Contributing

Issues and PRs are welcome. For adding new schemas or distribution channels, see CONTRIBUTING.md.

License

MIT — fork and customize freely. Derivative repos that don't merge back upstream (e.g. schema-invoice-custom) are also encouraged.