Roughdraft

A local-first markdown editor and viewer for working with AI.

{==Open one markdown file on your machine. Review it, comment on it, and suggest edits.==}{>>What does this mean?<<}{id="c3" by="user" at="2026-04-30T20:18:51.163Z"}{>>It means Roughdraft works with a normal local Markdown file: you open one .md file from your computer, read it in the app, leave inline comments, and propose edits that are saved back into the Markdown using CriticMarkup.<<}{id="c4" by="AI" at="2026-04-30T20:19:39.000Z" re="c3"}{>>cjool<<}{id="c5" by="user" at="2026-05-07T20:38:25.621Z" re="c4"}

Paste this into your coding agent:

Install Roughdraft for me using `npm i -g roughdraft`, then read https://roughdraft.page/setup.md and set yourself up to use it.

Or install and open a file yourself:

npm i -g roughdraft
roughdraft open /absolute/path/to/file.md

What is this?

Roughdraft is a local-first markdown editor and viewer that runs on your computer.

Its job is to make markdown files easy to open, read, edit, review, and discuss with your AI agent without moving them into a proprietary format or a hosted app.

Roughdraft opens a single markdown file directly for CriticMarkup comments and suggested changes.

How it works

• Local-first markdown editor — Open normal .md files from your machine and edit them directly

• Works with your AI agent — Tell your local agent to open a file in Roughdraft on your computer, then keep collaborating from there

• Comments & suggested changes — Use CriticMarkup for inline feedback, revisions, and review conversations

• Markdown files on disk — Everything stays as regular markdown files you can also edit in VS Code, Vim, Cursor, or anywhere else

• No cloud, no account, no telemetry — Runs entirely on your machine

Quick start

Install Roughdraft and start the local server:

npm i -g roughdraft
roughdraft start

roughdraft start runs Roughdraft in the background, reuses or chooses a free localhost port, writes server state to ~/.roughdraft/server.json, prints the active URL, and exits while the server keeps running.

Open a specific markdown file:

roughdraft open ./path/to/my-essay/draft.md

For scripts and agents that need a URL without launching a browser:

roughdraft open ./path/to/my-essay/draft.md --print-url
roughdraft status --json

Check or stop the background server:

roughdraft status
roughdraft stop

roughdraft open will reuse the running server and auto-start it if needed. You can also use roughdraft ./path/to/file.md as a shortcut when the input clearly looks like a path.

Roughdraft does not edit ~/CLAUDE.md, ~/AGENTS.md, or other user-level agent files. The setup prompt asks your agent to update its own guidance.

If the local server is already running, you can also open a file directly by URL:

http://localhost:7373/?path=/absolute/path/to/my-essay/draft.md

That makes an agent-friendly workflow possible:

1. Your AI writes or updates markdown files on disk.

2. You tell it to open a markdown file in Roughdraft.

3. Roughdraft opens locally on your machine.

4. You read, edit, leave comments, and suggest changes.

5. You click Done Reviewing in Roughdraft, and the AI can respond to your comments or revise the document.

Agents can watch that handoff directly:

roughdraft open ./path/to/my-essay/draft.md --json

roughdraft open starts or reuses the local server, opens the document, registers a fresh watcher, blocks until the next review.completed event, then prints event JSON with the document path, file version, and feedback counts. By default there is no watch timeout; pass --timeout <seconds> when you want one. Use --no-watch when you only want to open the document and return immediately. If no watcher is active when you click Done Reviewing, Roughdraft shows a fallback prompt you can copy into the agent.

Experimental MCP clients can start the stdio server with:

roughdraft mcp

The MCP server exposes tools to read the review index, list pending feedback, watch review events, append replies, and mark items resolved. CriticMarkup in the Markdown file remains the durable source of truth.

Local development

./scripts/setup.sh
./scripts/run.sh

./scripts/setup.sh installs workspace dependencies and builds the app and server. ./scripts/run.sh serves the built app at http://localhost:7373.

The two scripts coordinate through a lock file, so it's safe to start ./scripts/run.sh while ./scripts/setup.sh is still in progress. run will wait for setup to finish, or trigger setup itself if nothing has been built yet.

If you prefer package scripts, the same commands are available as pnpm setup and pnpm start.

Running pnpm setup also installs a per-worktree dev CLI wrapper into ~/.local/bin by default, using the current worktree directory name. For example, this checkout might install roughdraft-dev-lyon-v2, which points at this worktree's local code while leaving the published global roughdraft command untouched.

Each dev wrapper keeps its own server state under ~/.roughdraft/dev/<wrapper-name> by default, so opening a file from one worktree will not accidentally reuse a backend started from another worktree. roughdraft-dev-<worktree> open ... can start its own background server as needed; you do not need to run pnpm dev first just to open files in Roughdraft.

You can refresh that wrapper manually with:

pnpm dev:install-cli
pnpm dev:install-cli --name api-redesign

Quality checks:

pnpm lint
pnpm test
pnpm check

pnpm check is the same command the pull request workflow runs before merge.

Publishing

Roughdraft publishes from main when the root package.json version is newer than the current npm latest version.

Release flow:

1. Bump the root package.json version in a pull request.

2. Merge the pull request to main.

3. The Publish to npm GitHub Actions workflow runs pnpm check, publishes the package if that exact version is not already on npm and is newer than latest, then creates a v<version> git tag.

The workflow uses npm trusted publishing, so npm must be configured with this trusted publisher:

Owner: Lex-Inc
Repository: roughdraft
Workflow filename: publish.yml

No NPM_TOKEN secret is required.

Files on disk

my-essay/
  draft-1.md            # A normal markdown file on disk
  draft-2.md            # Another file you can open separately

Roughdraft reads and writes the markdown file directly.

Agent setup

If you want your local agent to remember the Roughdraft workflow, ask it to read the live setup prompt:

Install Roughdraft for me using `npm i -g roughdraft`, then read https://roughdraft.page/setup.md and set yourself up to use it.

Use roughdraft help, roughdraft help agent, or roughdraft help criticmarkup if you need a local refresher.

CLI reference

roughdraft [flags] <command> [args]
roughdraft <path>

Commands:

open <path>        Open one Markdown file and wait for Done Reviewing
start              Start or reuse the background server
status             Show server status
stop               Stop the managed background server
watch <path>       Wait for a Done Reviewing event
mcp                Start the experimental stdio MCP server
doctor [path]      Diagnose setup or validate Markdown
help agent         Print the agent setup prompt
help criticmarkup  Show CriticMarkup examples
agent-setup        Print the agent setup prompt
criticmarkup       Show CriticMarkup examples

Global flags:

-h, --help         Show help
--version          Print version
--json             Print JSON for supported commands
--no-color         Disable color

Useful command flags:

roughdraft open <path> --no-open
roughdraft open <path> --print-url
roughdraft open <path> --json
roughdraft open <path> --no-watch
roughdraft start --port <port>
roughdraft status --json
roughdraft stop --all
roughdraft watch ./draft.md --json
roughdraft doctor --json
roughdraft doctor ./draft.md
roughdraft doctor ./draft.md --json

Usage errors return exit code 2. Runtime failures return exit code 1. roughdraft status --json returns exit code 0 even when the JSON says "running": false.

Supported environment variables:

ROUGHDRAFT_PORT
  Preferred server port.

PORT
  Legacy preferred server port. Used only when ROUGHDRAFT_PORT is unset.

ROUGHDRAFT_NO_OPEN=1
  Disable browser/app opening.

ROUGHDRAFT_STATE_FILE
  Exact path to the server state JSON file.

ROUGHDRAFT_STATE_DIR
  Directory containing server.json.

Development-only environment variables:

ROUGHDRAFT_DEV_FRONTEND_STATE_FILE
ROUGHDRAFT_DEV_BIN_DIR
ROUGHDRAFT_DEV_STATE_BASE_DIR
ROUGHDRAFT_DEV_WRAPPER_NAME
ROUGHDRAFT_DEV_WRAPPER_PATH
ROUGHDRAFT_DEV_WRAPPER_REPO_ROOT

Roughdraft-flavored CriticMarkup

Roughdraft uses CriticMarkup as the readable review layer inside normal Markdown files. It supports the standard markers for comments, highlights, insertions, deletions, and substitutions:

The canonical Roughdraft Flavored Markdown spec is published at roughdraft.page/spec/roughdraft-flavored-markdown.md. The review-index JSON Schema is published at roughdraft.page/spec/roughdraft-flavored-markdown.schema.json.

This is {--deleted--} text.
This is {++inserted++} text.
This is {~~old~>new~~} substituted text.
This is {>>a comment<<} in the margin.
This is {==highlighted==} text.

Roughdraft extends those markers with compact attribute blocks so review state can round-trip through the file. Attribute blocks are written immediately after the comment or suggestion:

Please revisit {==this sentence==}{>>Needs a source<<}{id="c1" by="user" at="2026-04-28T12:00:00.000Z"}.

Supported attributes:

• id gives the comment or suggested change a stable document-local id.

• by records the reviewer or agent that created it.

• at records an ISO timestamp.

• re links a reply to another comment or suggestion id.

Replies are stored as additional comment blocks that point at the parent id:

Please revisit {==this sentence==}{>>Needs a source<<}{id="c1" by="user" at="2026-04-28T12:00:00.000Z"}{>>I can add one from the intro.<<}{id="c2" by="AI" at="2026-04-28T12:05:00.000Z" re="c1"}.

Suggested changes can also carry ids and discussion:

Add {++one concrete example++}{id="s1" by="AI" at="2026-04-28T12:10:00.000Z"}{>>Use the customer story here.<<}{id="c3" by="user" at="2026-04-28T12:12:00.000Z" re="s1"}.
Remove {--vague phrasing--}{id="s2" by="user" at="2026-04-28T12:13:00.000Z"}.
Use {~~rough~>specific~~}{id="s3" by="AI" at="2026-04-28T12:14:00.000Z"} wording.

CriticMarkup inside inline code and fenced code blocks is treated as literal example text, not live review feedback:

Inline code stays literal: `{==not a comment==}`.

```text
{++not a suggestion++}
```

This matters because the main workflow is often:

• The AI writes a doc

• The user opens it in Roughdraft

• The user leaves comments and suggested changes

• The AI reads those comments and responds in the same markdown file

Try the demo

Don't want to install anything? Try the live demo — it runs entirely in your browser using local storage.

License

MIT

Built by Nathan Baschez