Unlock the full potential of your reMarkable tablet as a second brain for AI assistants. This MCP server lets Claude, VS Code Copilot, and other AI tools read, search, and traverse your entire reMarkable library — including handwritten notes via OCR.
Your reMarkable tablet is a powerful tool for thinking, note-taking, and research. But that knowledge stays trapped on the device. This MCP server changes that:
- Full library access — Browse folders, search documents, read any file
- Typed text extraction — Native support for Type Folio and typed annotations
- Handwriting OCR — Convert handwritten notes to searchable text
- PDF & EPUB support — Extract text from documents, plus your annotations
- Robust page rendering — Renders pages locally and automatically falls back to a source PDF when the local stroke renderer can't (USB/SSH use the tablet's own PDF export; cloud uses the original source PDF), so images work across firmware versions and even without system graphics libraries installed
- Smart search — Find content across your entire library
- Second brain integration — Use with Obsidian, note-taking apps, or any AI workflow
Whether you're researching, writing, or developing ideas, remarkable-mcp lets you leverage everything on your reMarkable through AI.
Connect via USB and enable the web interface in your tablet's Storage Settings.
Setup:
- Connect your reMarkable via USB
- On your tablet: Settings → Storage → Enable "USB web interface"
- Install via the button above
Why USB Web?
- ✅ Fast offline access over USB
- ✅ No subscription required
- ✅ Simple — just enable in Storage Settings
📋 Manual USB Web Configuration
Add to .vscode/mcp.json:
{
"servers": {
"remarkable": {
"command": "uvx",
"args": ["remarkable-mcp", "--usb"],
"env": {
"GOOGLE_VISION_API_KEY": "your-api-key"
}
}
}
}Troubleshooting:
- Make sure your reMarkable is connected via USB and unlocked
- Verify USB web interface is enabled in Settings → Storage
- The tablet should be accessible at
http://10.11.99.1
For power users who need direct filesystem access. Faster than USB Web but requires developer mode (factory reset).
Requirements: Developer mode enabled + USB connection to your reMarkable
📋 Manual SSH Configuration
Add to .vscode/mcp.json:
{
"servers": {
"remarkable": {
"command": "uvx",
"args": ["remarkable-mcp", "--ssh"],
"env": {
"GOOGLE_VISION_API_KEY": "your-api-key"
}
}
}
}See SSH Setup Guide for detailed instructions.
Wireless access with no device connection required — your reMarkable syncs to the cloud and the MCP reads from there, so it works from anywhere. Requires a reMarkable Connect subscription.
Cloud mode fetches your whole library in parallel and caches content-addressed blobs on disk, so after the first run startups and document reads are near-instant (a 388-document library lists in ~4s cold, ~0.5s warm). See Cloud Performance & Caching to tune it.
📋 Cloud Mode Setup
Go to my.remarkable.com/device/desktop/connect and generate a code.
uvx remarkable-mcp --register YOUR_CODEOr configure manually in .vscode/mcp.json:
{
"inputs": [
{
"type": "promptString",
"id": "remarkable-token",
"description": "reMarkable API Token",
"password": true
},
{
"type": "promptString",
"id": "google-vision-key",
"description": "Google Vision API Key",
"password": true
}
],
"servers": {
"remarkable": {
"command": "uvx",
"args": ["remarkable-mcp"],
"env": {
"REMARKABLE_TOKEN": "${input:remarkable-token}",
"GOOGLE_VISION_API_KEY": "${input:google-vision-key}"
}
}
}
}All three modes share the same read, render, and upload tools. Cloud and SSH additionally support full library management — create folders, move, rename, and delete (all enabled by default) — so capability is near-identical and you can genuinely pick whichever matches how your tablet is connected:
- ☁️ Cloud — device-free, works from anywhere. Reads your library straight from reMarkable's cloud over Wi‑Fi with a Connect subscription — no cable, no developer mode. Full read/render plus full write (upload, create folder, move, rename, delete → trash). Parallel fetching and an on-disk blob cache make it fast after the first sync. Best for remote/headless setups or when you don't want to plug in.
- 🔌 USB Web Interface — best when the tablet is plugged in. Enable the web interface in Storage Settings — no subscription, no developer mode. Full read/render plus upload (to your root folder). The tablet's USB web firmware exposes no folder/move/rename/delete endpoints, so for those over a cable use SSH.
- ⚡ SSH — for power users who want filesystem-level access. Requires developer mode over USB. Full read/render plus full write including folder create/move/rename/delete, straight from the tablet filesystem.
| Mode | Setup | Subscription | Offline | Read + render | Raw PDF/EPUB | Upload | Folder ops¹ |
|---|---|---|---|---|---|---|---|
| ☁️ Cloud | One-time code | Connect | ❌ | ✅ | ✅ PDF/EPUB | ✅ | ✅ |
| 🔌 USB Web | Enable in Settings | Not required | ✅ | ✅ | ✅ (to root) | ❌ | |
| ⚡ SSH | Developer mode | Not required | ✅ | ✅ | ✅ PDF/EPUB | ✅ | ✅ |
¹ Folder ops = create folder / move / rename / delete. Upload and folder ops are enabled by default; pass --read-only to expose a read-only server. Deletes move items to the trash and prompt for confirmation when your client supports elicitation, and are refused without it unless REMARKABLE_SKIP_CONFIRM=1 is set.
If you select a device transport (--usb or --ssh) but the tablet isn't reachable at startup and a cloud token is configured (REMARKABLE_TOKEN or ~/.rmapi), the server automatically falls back to cloud mode and logs a warning. This means a single configuration works whether or not the tablet is plugged in — plug in for fast local access, unplug to keep working over the cloud. remarkable_status reports the effective transport and a fell_back_to_cloud flag when this happens.
Pass --no-cloud-fallback (or set REMARKABLE_DISABLE_CLOUD_FALLBACK=1) to disable this and fail instead when the device is unreachable.
📖 Detailed Setup Guides:
- USB Web Interface Setup — recommended — simple setup, full feature support
- SSH Setup Guide — for advanced users who need filesystem access
- Cloud setup is documented in the Quick Install section above; tuning in Cloud Performance & Caching
remarkable-mcp works as an OpenClaw skill. Add to your openclaw.json:
{
"mcpServers": {
"remarkable": {
"command": "uvx",
"args": ["remarkable-mcp", "--usb"]
}
}
}Install from ClawHub:
clawhub install remarkable-mcpOr copy the SKILL.md from this repository into your ~/.openclaw/skills/remarkable-mcp/ directory.
| Tool | Description |
|---|---|
remarkable_read |
Read and extract text from documents (with pagination and search) |
remarkable_browse |
Navigate folders, search by document name, or filter by tags |
remarkable_search |
Search content across multiple documents (with tag filtering) |
remarkable_recent |
Get recently modified documents |
remarkable_status |
Check connection status and the per-transport capability matrix |
remarkable_image |
Get PNG/SVG images of pages (supports OCR via sampling) |
These six tools are read-only and return structured JSON with hints for next actions. Write tools (remarkable_upload, remarkable_mkdir, remarkable_move, remarkable_rename, remarkable_delete, and remarkable_author for native ink/notebooks) are enabled by default — pass --read-only to disable them — see Write Tools. An interactive canvas app (remarkable_canvas) is also registered automatically for clients that support MCP Apps.
- Auto-redirect — Browsing a document path returns its content automatically
- Auto-OCR — Notebooks with no typed text automatically enable OCR
- Batch search — Search across multiple documents in one call
- Vision support — Get page images for visual context (diagrams, mockups, sketches)
- Sampling OCR — Use client's AI for OCR on images (no API key needed)
- Tag support — Filter and organize documents by tags
# Read a document
remarkable_read("Meeting Notes")
# Search for keywords
remarkable_read("Project Plan", grep="deadline")
# Enable OCR for handwritten notes
remarkable_read("Journal", include_ocr=True)
# Browse your library
remarkable_browse("/Work/Projects")
# Filter by tags
remarkable_browse("/", tags=["important"])
remarkable_browse("/Work", tags=["project", "active"])
# Search across documents
remarkable_search("meeting", grep="action items")
# Search with tag filter
remarkable_search("project", tags=["work"])
# Get recent documents
remarkable_recent(limit=10)
# Get a page image (for visual content like UI mockups or diagrams)
remarkable_image("UI Mockup", page=1)
# Get SVG for editing in design tools
remarkable_image("Wireframe", output_format="svg")
# Get image with OCR text extraction (uses sampling if configured)
remarkable_image("Handwritten Notes", include_ocr=True)
# Transparent background for compositing
remarkable_image("Logo Sketch", background="#00000000")
# Compatibility mode: return resource URI instead of embedded resource
remarkable_image("Diagram", compatibility=True)Note: PNG rendering automatically falls back to a source PDF when the local stroke renderer can't produce an image (empty pages, newer
.rmformats, or a machine withoutlibcairo). USB and SSH modes use the tablet's native PDF export; cloud mode uses the document's original source PDF. This keepsremarkable_imageworking across firmware versions and platforms. Cloud mode has no native export, so it relies on the local renderer.
Documents are automatically registered as MCP resources:
| URI Scheme | Description |
|---|---|
remarkable:///{path}.txt |
Extracted text content |
remarkableraw:///{path}.pdf |
Original PDF file (SSH only) |
remarkableraw:///{path}.epub |
Original EPUB file (SSH only) |
remarkableimg:///{path}.page-{N}.png |
PNG image of page N (notebooks only) |
remarkablesvg:///{path}.page-{N}.svg |
SVG vector image of page N (notebooks only) |
📖 Full Resources Documentation
For handwritten content, remarkable-mcp offers several OCR backends. Choose based on your setup and requirements:
| Backend | Setup | Quality | Offline | Best For |
|---|---|---|---|---|
| Sampling | No API key | Depends on client model | ✅ | Users with capable AI clients |
| Google Vision | API key | Excellent | ❌ | Best handwriting accuracy |
| Tesseract | System install | Poor for handwriting | ✅ | Printed text, offline fallback |
Set REMARKABLE_OCR_BACKEND in your MCP config:
{
"env": {
"REMARKABLE_OCR_BACKEND": "sampling"
}
}Options: sampling, google, tesseract, auto
📖 Sampling OCR (No API Key)
Uses your MCP client's AI model for OCR. Works with clients that support MCP sampling (VS Code + Copilot, Claude Desktop, etc.).
Pros:
- No additional API keys needed
- Quality depends on your client's model (GPT-4, Claude, etc.)
- Private — handwriting stays local to your client
Cons:
- Only available with sampling-capable clients
- Falls back to Google Vision (if API key configured) or Tesseract if sampling unavailable
📖 Google Cloud Vision
Provides consistently excellent handwriting recognition.
Setup:
- Enable Cloud Vision API
- Create an API key
- Add to config:
"GOOGLE_VISION_API_KEY": "your-key"
Cost: 1,000 free requests/month, then ~$1.50 per 1,000.
📖 Tesseract (Fallback)
Open-source OCR designed for printed text. Poor results with handwriting, but useful as an offline fallback.
# Install Tesseract
# macOS
brew install tesseract
# Ubuntu/Debian
sudo apt install tesseract-ocr
# Windows
choco install tesseractWhen REMARKABLE_OCR_BACKEND=auto (default):
- Google Vision (if
GOOGLE_VISION_API_KEYis set) - Tesseract (fallback)
| Feature | SSH Mode | USB Web | Cloud API |
|---|---|---|---|
| Speed | ⚡ 10-100x faster | ⚡ Fast | ⚡ Fast (parallel + cached) |
| Offline | ✅ Yes | ✅ Yes | ❌ No |
| Subscription | ✅ Not required | ✅ Not required | ❌ Connect required |
| Raw files | ✅ PDFs, EPUBs | ✅ PDFs | ✅ PDFs, EPUBs |
| Upload | ✅ (default) | ✅ (default) | ✅ (default) |
| mkdir/move/rename/delete | ✅ (default) | ❌ | ✅ (default) |
| Setup | Developer mode | Enable in Settings | One-time code |
Write tools let you upload, organize, and manage documents on your reMarkable. Enabled by default. Cloud and SSH modes support the full set; USB web supports upload only (its firmware exposes no folder operations). Pass --read-only to expose a read-only server.
| Feature | Cloud Mode | SSH Mode | USB Web Mode |
|---|---|---|---|
| Upload | ✅ | ✅ | ✅ (to root) |
| Mkdir | ✅ | ✅ | ❌ |
| Move | ✅ | ✅ | ❌ |
| Rename | ✅ | ✅ | ❌ |
| Delete | ✅ (→ trash) | ✅ | ❌ |
Write tools are on by default in every mode. To run a read-only server, add the --read-only flag:
{
"servers": {
"remarkable": {
"command": "uvx",
"args": ["remarkable-mcp", "--read-only"]
}
}
}It combines with any transport flag (--ssh, --usb):
{
"servers": {
"remarkable": {
"command": "uvx",
"args": ["remarkable-mcp", "--ssh", "--read-only"]
}
}
}Or set the environment variable:
{
"env": {
"REMARKABLE_READ_ONLY": "1"
}
}The legacy
--writeflag andREMARKABLE_ENABLE_WRITEvariable are still accepted for backward compatibility but are now no-ops (write is the default).--writeand--read-onlyare mutually exclusive.
| Tool | Description |
|---|---|
remarkable_upload(file_path, parent_folder, document_name) |
Upload a PDF or EPUB file (all modes; USB web ignores folder/name and uploads to root) |
remarkable_mkdir(folder_name, parent) |
Create a new folder (cloud and SSH) |
remarkable_move(document, dest_folder) |
Move a document or folder (cloud and SSH) |
remarkable_rename(document, new_name) |
Rename a document or folder (cloud and SSH) |
remarkable_delete(document) |
Delete a document or folder — destructive (cloud and SSH) |
remarkable_author(method, ...) |
Author native ink and notebooks — draw (append strokes), add_page (append a blank notebook page), create_document (new notebook) — SSH only |
- Upload registers in all modes — cloud, SSH, and USB web.
- mkdir, move, rename, delete register in cloud and SSH modes only — they are not exposed on USB web (the tablet's USB web firmware has no folder/move/rename/delete endpoints), keeping the tool list scoped to what the active transport actually supports.
- Delete prompts for confirmation when possible — if the client supports MCP elicitation,
remarkable_deleteasks the user to confirm before deleting. If the client can't show a prompt, the delete is refused (not performed) unlessREMARKABLE_SKIP_CONFIRM=1is set — so write-on-by-default can't silently delete from clients that lack elicitation. In cloud mode delete moves the item to the trash (recoverable from your device); setREMARKABLE_SKIP_CONFIRM=1to allow deletes without a prompt in automated setups. All write tools carryToolAnnotations(readOnlyHint=False)(anddestructiveHint=Truefor delete) so an agent harness can gate writes at the MCP layer. - After each write operation in SSH mode, the tablet UI restarts automatically to reflect changes.
# Upload a PDF
remarkable_upload("paper.pdf", parent_folder="/Research")
# Create a folder
remarkable_mkdir("2024 Archive", parent="/Archive")
# Move a document
remarkable_move("Meeting Notes", "/Archive/2024 Archive")
# Rename a document
remarkable_rename("Untitled", "Q4 Planning Notes")
# Delete (destructive — confirms via elicitation when supported)
remarkable_delete("Old Draft")
# Author native ink and notebooks (SSH only)
# Append pen/highlighter strokes to a page (coordinates normalized [0,1] from
# the page's top-left). The interactive canvas Save button calls this too.
remarkable_author(
method="draw", document="Ideas", page=1,
strokes=[{"points": [[0.1, 0.2], [0.8, 0.2]], "tool": "highlighter", "color": "yellow"}],
)
# Append a blank, drawable page to the end of a notebook
remarkable_author(method="add_page", document="Ideas")
# Create a new (blank) notebook — the common case
remarkable_author(method="create_document", name="Sketches")
# Only seed typed text when the user explicitly requested it.
remarkable_author(method="create_document", name="Meeting notes", text="Agenda\nFollow-ups")An interactive page viewer built on the MCP Apps extension (SEP-1865). Clients that support MCP Apps (such as ChatGPT, Claude, VS Code, and the MCP Inspector) render a canvas in a side panel where you can view a document page and navigate through it.
There is no flag to enable it — the remarkable_canvas tool and its ui://remarkable/canvas resource are always registered, and the capability is negotiated automatically at the MCP initialize handshake. App-capable clients open the interactive canvas; every other client simply receives the rendered page as an image, so the tool is safe and useful everywhere.
This registers one tool:
| Tool | Description |
|---|---|
remarkable_canvas(document, page) |
Open a page in the interactive canvas viewer |
How it behaves:
- App-capable clients open the canvas (declared at
ui://remarkable/canvas, MIMEtext/html;profile=mcp-app) and can page through the document via the MCP Apps postMessage bridge — the server delivers each rendered page in the tool result'sstructuredContent. - Other clients still get the rendered page back as an embedded PNG image, so the tool is useful everywhere; it just won't open the interactive panel. The
_meta.ui/ui://metadata is inert to clients that don't advertise the MCP Apps UI extension.
When write mode is on (the default) and the active transport is SSH, the canvas becomes a write surface:
- Draw — pick a pen or highlighter and colour, draw over the page, and Save writes the strokes back to the device as native
.rmink. Strokes are buffered locally per page (with Undo and Cancel) and only touch the device on Save. - + Page (native notebooks only) — queues a new blank page locally that you can navigate to and draw on immediately. Save materializes the queued page(s) on the device first, then writes any cached strokes.
- One source of truth: the canvas calls the same
remarkable_authortool a model would call (method="draw"on Save,method="add_page"for +Page), so the human path and the model path produce byte-identical results.
The Save / Draw / +Page controls are hidden when the page isn't writable (read-only mode, or a non-SSH transport), and the canvas falls back to a plain image viewer. The iframe bridge follows the MCP Apps spec but is best validated against your specific client.
Limit the MCP server to a specific folder on your reMarkable. All operations will be scoped to this folder:
{
"servers": {
"remarkable": {
"command": "uvx",
"args": ["remarkable-mcp", "--ssh"],
"env": {
"REMARKABLE_ROOT_PATH": "/Work",
"GOOGLE_VISION_API_KEY": "your-api-key"
}
}
}
}With this configuration:
remarkable_browse("/")shows contents of/Workremarkable_browse("/Projects")shows/Work/Projects- Documents outside
/Workare not accessible
Useful for:
- Focusing on work documents during office hours
- Separating personal and professional notes
- Limiting scope for specific AI workflows
Set the default background color for image rendering:
{
"servers": {
"remarkable": {
"command": "uvx",
"args": ["remarkable-mcp", "--ssh"],
"env": {
"REMARKABLE_BACKGROUND_COLOR": "#FFFFFF"
}
}
}
}Supported formats:
#RRGGBB— RGB hex (e.g.,#FFFFFFfor white)#RRGGBBAA— RGBA hex (e.g.,#00000000for transparent)
Default is #FBFBFB (reMarkable paper color). This affects both the remarkable_image tool and image resources.
Cloud API requests automatically retry on transient failures (HTTP 429, 500, 502, 503, 504) and network errors with exponential backoff and jitter. You can tune this via environment variables:
| Variable | Default | Description |
|---|---|---|
REMARKABLE_RETRY_ATTEMPTS |
3 |
Maximum number of request attempts (minimum 1) |
REMARKABLE_RETRY_DELAY |
2.0 |
Base delay in seconds for exponential backoff |
The retry logic honours the Retry-After header from rate-limited responses — both the numeric (seconds) form and the HTTP-date form (which Cloudflare, fronting the reMarkable cloud, often sends) — capped at 20 seconds. Auth failures (401) are not retried — they trigger automatic token renewal instead.
Cloud mode is built to make a device-free workflow fast:
- Parallel traversal — document metadata is fetched concurrently instead of one document at a time, turning a multi-minute first load into a few seconds.
- Connection pooling — HTTP connections are reused (keep-alive), avoiding a fresh TLS handshake per request.
- Content-addressed blob cache — reMarkable's cloud is an immutable, hash-addressed store (like Git), so a blob's bytes can never change for a given hash. Downloaded blobs are cached on disk and reused on later runs; changed documents get new hashes and are re-fetched automatically. This makes warm startups and repeat document reads near-instant, and it is invalidation-safe by construction.
You normally don't need to configure any of this, but these environment variables let you tune it:
| Variable | Default | Description |
|---|---|---|
REMARKABLE_SYNC_WORKERS |
16 |
Parallel workers for cloud fetches (clamped to 64). |
REMARKABLE_DISABLE_CACHE |
unset | Set to 1 to disable the on-disk blob cache entirely. |
REMARKABLE_CACHE_DIR |
~/.remarkable/cache/blobs |
Where cached blobs are stored. |
REMARKABLE_CACHE_MAX_BLOB |
4194304 (4 MiB) |
Blobs larger than this are streamed through but not cached. |
The cache is purely a local accelerator: deleting REMARKABLE_CACHE_DIR only forces the next read to re-download. The mutable cloud root hash is always fetched fresh, so you never see a stale library.
Use remarkable-mcp while working in an Obsidian vault or similar to transfer knowledge from your handwritten notes into structured documents. AI can read your research notes and help develop your ideas.
Ask your AI assistant to summarize your recent notes, find action items, or identify patterns across your journal entries.
Find that half-remembered note by searching across your entire library — including handwritten content.
Treat your reMarkable as a second brain that AI can access. Combined with tools like Obsidian, you can build a powerful personal knowledge system.
| Guide | Description |
|---|---|
| SSH Setup | Enable developer mode and configure SSH |
| Google Vision Setup | Set up handwriting OCR |
| Tools Reference | Detailed tool documentation |
| Resources Reference | MCP resources documentation |
| Capability Negotiation | MCP protocol capabilities |
| Development | Contributing and development setup |
| Future Plans | Roadmap and planned features |
git clone https://github.com/SamMorrowDrums/remarkable-mcp.git
cd remarkable-mcp
uv sync --all-extras
uv run pytest test_server.py -vWhen something looks broken, run the deterministic, no-AI smoke test first. It drives the real server over MCP and exercises every available tool in every reachable transport (cloud → usb-web → ssh):
uv run python smoke/run_smoke.py # all available modes
uv run python smoke/run_smoke.py --read-only # connectivity + reads only📖 smoke/README.md — what PASS / N/A / SKIP / FAIL mean and per-mode expectations.
MIT
Built with rmscene, PyMuPDF, and inspiration from ddvk/rmapi.