An open API service indexing awesome lists of open source software.

https://github.com/travsteward/openwriter

The open-source writing surface for AI agents. Markdown-native editor with pending change review — your agent writes, you accept or reject.
https://github.com/travsteward/openwriter

ai ai-agents ai-writing anthropic claude claude-code claude-skills collaborative-editing content-creation editor markdown mcp model-context-protocol newsletter open-source publishing tiptap writing writing-tools

Last synced: 24 days ago
JSON representation

The open-source writing surface for AI agents. Markdown-native editor with pending change review — your agent writes, you accept or reject.

Awesome Lists containing this project

README

          

# OpenWriter

**The open-source writing surface for the agentic era.**

A markdown-native rich text editor built for human-agent collaboration. Your AI agent writes, you review. Plain `.md` files on disk — no database, no lock-in. Works with any MCP-compatible agent.

![OpenWriter — agent writes, you review](assets/screenshot.png)

---

## Why OpenWriter?

Claude Code, OpenCode, Codex — these agents are transforming how people build software. But that power hasn't crossed over to writing. Ask an agent to edit your essay and you're staring at raw markdown in a terminal. There's no collaborative surface, no way to review changes, no real workflow.

Every writing tool is racing to bolt on its own weak agent. They have it backwards. The most powerful agents already exist — and they're designed to work with tools, not be trapped inside one. OpenWriter doesn't ship an agent. It's the first writing surface built for the agents you already use, with a review system that keeps you in control.

Markdown is the native language of AI. Every LLM reads it, writes it, and thinks in it. OpenWriter treats `.md` files as first-class citizens — your documents are plain markdown on disk, and the editor adds rich formatting, workspaces, version history, and agent collaboration on top. No proprietary format. No database. Just files.

**The agent writes. You accept or reject. That's it.**

- Documents are plain `.md` files — open existing ones or create new ones
- Agent makes changes → they appear as colored decorations (green for inserts, blue for rewrites, red for deletions)
- You review with vim-style hotkeys (`j`/`k` navigate, `a` accept, `r` reject)
- Cross-document navigation when an agent edits multiple files at once
- Works with any MCP agent — no vendor lock-in

---

## Quick Start

```bash
npm install -g openwriter
openwriter
```

That's it. Opens your browser to `localhost:5050` with a ready-to-use editor. Documents save as markdown files in `~/.openwriter/`.

> **Try without installing:** `npx openwriter` works too, but global install starts instantly (no npm resolution delay on every launch).

Already have markdown files? Open them directly — the agent can use `open_file` to load any `.md` from disk, or you can drag files into the sidebar.

### Connect Your Agent

**Install the skill** (Claude Code, Cursor, Codex, and 20+ agents):
```bash
npx skills add https://github.com/travsteward/openwriter --skill openwriter
```

Then add the MCP server for the 24 editing tools:
```bash
claude mcp add -s user openwriter -- openwriter --no-open
```

The skill teaches your agent how to use OpenWriter's tools effectively — writing strategy, review etiquette, and troubleshooting. The MCP server provides the actual document editing capabilities.

**Other MCP agents** (Cursor, OpenCode, etc.) — add to your MCP config:

```json
{
"mcpServers": {
"openwriter": {
"command": "openwriter",
"args": ["--no-open"]
}
}
}
```

Now your agent has 24 tools to read, write, and organize documents — and every change goes through your review.

---

## Features

### Agent Collaboration via MCP

24 tools across four categories:

| Category | Tools | What They Do |
|----------|-------|-------------|
| **Document** | `read_pad`, `write_to_pad`, `edit_text`, `get_pad_status`, + 5 more | Read/write document content, fine-grained text edits, metadata |
| **Multi-doc** | `list_documents`, `switch_document`, `create_document` | Navigate and manage multiple documents |
| **Workspace** | `create_workspace`, `get_workspace_structure`, `add_doc`, + 6 more | Organize docs into projects with containers and tags |
| **Import** | `import_gdoc` | Import structured Google Docs, auto-split into chapters |

Agents write in markdown or TipTap JSON. The server converts, assigns node IDs, and broadcasts changes to your browser in real-time via WebSocket.

### Pending Change Review

The core interaction model. When an agent (or the context menu) makes changes:

- **Inserts** appear highlighted in green
- **Rewrites** appear highlighted in blue (original content preserved for reject)
- **Deletions** appear with red strikethrough

Review Panel (floating bottom bar):

| Key | Action |
|-----|--------|
| `j` / `k` | Next / previous change |
| `h` / `l` | Previous / next document with changes |
| `a` | Accept current change |
| `r` | Reject current change |
| `Shift+A` | Accept all in document |
| `Shift+R` | Reject all in document |

### Multi-Document Workspaces

Documents are markdown files on disk. Organize them into workspaces with nested containers, cross-cutting tags, and shared context (characters, settings, rules) that agents can read for consistency.

Four sidebar views:
- **Tree** — Hierarchical folders with drag-and-drop
- **Timeline** — Sorted by last modified
- **Board** — Card-based drill-down navigation
- **Shelf** — Visual bookshelf metaphor with spine browsing

### Context Menu (Right-Click)

Select text and right-click for AI-powered transformations:

| Action | Key | Description |
|--------|-----|-------------|
| Rewrite | `R` | Rewrite selection at similar length |
| Shrink | `S` | Condense by 40-60% |
| Expand | `E` | Expand by 50-100% |
| Custom | — | Free-text instruction |
| Fill | `F` | Generate content between paragraphs |
| Insert after | `I` | Generate new content after selection |
| Delete | `D` | Mark for deletion |
| Link to doc | `L` | Create internal document links |

Context menu actions are provided by plugins. The built-in [Author's Voice](https://authors-voice.com) plugin rewrites text in your personal writing voice.

### Themes

5 themes, each with light and dark modes:

- **Ink** — Clean, minimal, professional
- **Novel** — Warm, serif-based, literary
- **Mono** — Monospace, code-focused
- **Editorial** — Bold magazine-style headings
- **Studio** — Contemporary sans-serif

Three typography presets (default, compact, expanded) work with any theme.

### Git Sync

Push your documents to GitHub directly from the editor. Three setup methods:
- **GitHub CLI** — Auto-detected if `gh` is authenticated
- **Personal Access Token** — Manual GitHub auth
- **Existing repo** — Connect to a repo you already have

### Export

Export any document to:
- Markdown (`.md`)
- HTML (styled web page)
- Word (`.docx`)
- Plain text (`.txt`)
- PDF (via print preview)

### Version History

Automatic snapshots with full rollback. Browse previous versions and restore any point.

---

## Markdown Native

Every document is a `.md` file on disk. What you see in the editor is markdown with rich rendering — headings, lists, tables, code blocks, images, links — all stored as plain text.

```
~/.openwriter/
├── Getting Started.md
├── Chapter 1 - Origins.md
├── Research Notes.md
└── _workspaces/
└── My Novel.json
```

- **No database.** The filesystem is the index. Move, copy, or `grep` your files however you want.
- **Open any `.md` file.** Point OpenWriter at existing markdown from any project — it loads instantly.
- **Git Sync built in.** Push your documents to GitHub directly from the editor. Your markdown files are version-controlled and portable — access them from any machine or future web client.
- **Frontmatter metadata.** YAML frontmatter for tags, status, or any key-value pairs your workflow needs.
- **Full markdown fidelity.** Bold, italic, strikethrough, code blocks with syntax highlighting, tables, task lists, images, links, subscript, superscript — all round-trip cleanly to `.md`.
- **AI-native format.** Every LLM reads and writes markdown natively. No conversion layer, no token waste. The agent edits the same format the file is stored in.

---

## Token-Efficient Wire Format

Agents don't parse JSON. OpenWriter uses a compact tagged-line format that's ~10x more token-efficient:

```
title: My Document
words: 1,205
pending: 2
---
[h1:a1b2c3d4] Chapter One
[p:e5f6g7h8] The quick brown fox jumped over the **lazy** dog.
[ul:i9j0k1l2]
[li:m3n4o5p6] First bullet
[li:q7r8s9t0] Second bullet
```

Each line: `[type:8-char-id] content` with inline markdown preserved. Agents read and write naturally.

---

## Plugin System

OpenWriter is extensible via plugins. A plugin can:

- **Register MCP tools** — Extend the agent's capabilities
- **Add HTTP routes** — Custom API endpoints on the server
- **Contribute context menu items** — UI actions for text transformation

```typescript
import type { OpenWriterPlugin } from 'openwriter';

const plugin: OpenWriterPlugin = {
name: 'my-plugin',
version: '1.0.0',

mcpTools(config) {
return [{
name: 'my-tool',
description: 'Does something useful',
inputSchema: { type: 'object', properties: {} },
handler: async (params) => ({ result: 'done' })
}];
},

contextMenuItems() {
return [{
label: 'My Action',
action: 'myplugin:do-thing',
condition: 'has-selection'
}];
}
};

export default plugin;
```

Load plugins at startup:

```bash
openwriter --plugins my-plugin,another-plugin
```

---

## CLI Options

```bash
openwriter [options]

Options:
--port Port number (default: 5050)
--no-open Don't auto-open browser
--api-key Author's Voice API key
--av-url Author's Voice backend URL
--plugins Comma-separated plugin names

Subcommands:
install-skill Install Claude Code companion skill to ~/.claude/skills/openwriter/
```

Environment variables: `AV_API_KEY`, `AV_BACKEND_URL`

---

## Architecture

```
Browser (localhost:5050)
├── TipTap 3.0 Editor (React)
├── Decoration Plugin (pending insert/rewrite/delete)
├── Review Panel (accept/reject with keyboard nav)
├── Sidebar (4 views: tree, timeline, board, shelf)
└── Context Menu (plugin-provided AI actions)

│ WebSocket + HTTP

Pad Server (Express + WebSocket + MCP stdio)
├── Document state (in-memory + markdown on disk)
├── 24 MCP tools + plugin tools
├── Workspace management
├── Git sync, versions, export
└── Plugin loader

│ MCP stdio

AI Agent (Claude Code, Cursor, etc.)
```

Three interfaces:
- **HTTP** — Browser UI operations, document CRUD, plugin proxying
- **WebSocket** — Real-time push of agent changes to browser
- **MCP stdio** — Agent reads/writes documents

The server supports **multi-session mode**: if port 5050 is already taken, additional instances proxy MCP calls via HTTP to the running server. Multiple agents can safely share the same document state.

---

## Development

```bash
# Clone and install
git clone https://github.com/travsteward/openwriter.git
cd openwriter
npm install

# Dev mode (hot reload)
cd packages/openwriter
npm run dev

# Build
npx turbo run build --force

# Type check
npx tsc --noEmit -p packages/openwriter/tsconfig.json # frontend
npx tsc --noEmit -p packages/openwriter/tsconfig.server.json # server

# Run production build
node packages/openwriter/dist/bin/pad.js
```

Monorepo structure: `packages/openwriter` (editor + server), `plugins/` (optional extensions).

---

## Contributing

Contributions welcome. Please open an issue first to discuss what you'd like to change.

---

## License

[MIT](LICENSE) — Travis Steward