Open Private Cloud for Supernote
A lightweight, self-hosted Python private cloud server, binary .note parser & authoring library, and optional AI intelligence layer for your Ratta Supernote.
Unofficial Community Project: We love Ratta's products and built this toolkit to support the Supernote community. It is an independent, open-source project and is not affiliated with, endorsed by, or sponsored by Shanghai Ratta Smart Technology Co., Ltd. Supernote and Ratta are trademarks of Shanghai Ratta Smart Technology Co., Ltd.
supernote is an open-source Python toolkit for Supernote hardware (Nomad, A5 X2 Manta, A5 X, A6 X) built around three core components:
- A Lightweight Private Cloud Server (
supernote.server): A single-processasyncio+ SQLite server that implements 100% of the community Supernote OpenAPI Specification in ~135 MB of RAM (bare sync mode) or ~160 MB idle / ~310 MB peak when rendering and processing notebooks with the Web UI, Model Context Protocol (MCP) server, and optional AI synthesis engine. - A Standalone
.noteParser, Converter & Authoring Library (supernote.notebook): Read binary.noteand.markfiles, export pages to PDF (with clickable links), PNG, vector SVG, or plain text, or author and append Bullet Journal spreads directly into.notefiles from Markdown. - An Async Python Client SDK & CLI (
supernote.client,supernote.cli): Automate file sync, task lists, semantic search, static Bearer API keys (sn_...), and real-time Socket.IO sync notifications from Python or the terminal.
Why Open Private Cloud for Supernote?
Native Tablet Sync in ~135 MB of RAM (No Java, MariaDB, or Redis)
While the official Supernote Private Cloud stack orchestrates multi-container Java/Spring Boot, MariaDB, and Redis services (~2 GB+ RAM), Open Private Cloud runs in a single Python asyncio process backed by a single SQLite file and content-addressable blob storage. Point your tablet's native Settings → Sync → Private Cloud to your server and tap Sync—it runs effortlessly on a Raspberry Pi, NAS, or low-memory home lab server with two-tier volume and per-user storage quotas.
Incremental Handwriting OCR, Summaries & Semantic Search
Handwriting recognition on e-ink hardware drains battery and struggles with dense cursive, diagrams, and Bullet Journal signifiers. When a notebook syncs to Open Private Cloud, a background pipeline hashes every page (PageHashingModule), skips unchanged and blank pages automatically, and processes only modified pages using Vision LLMs (Google Gemini gemini-3.5-flash-lite by default, or local models via Ollama / vLLM / llama.cpp):
- High-Fidelity Transcripts: Converts handwritten pages, tables, diagrams, and Bullet Journal Rapid Logging symbols (
•,X,>,<,-,*,~~strikethroughs~~) into structured Markdown. - Executive Summaries: Synthesizes multi-page notebooks into scannable summaries with key takeaways.
- Concept Search: Vectorizes content so you can search across years of handwritten notebooks by meaning (
"Where did I sketch the backyard drainage plan?"), not just filenames.
Built-In MCP Server & Static Bearer API Keys (sn_...)
Out of the box, the server hosts a spec-compliant Model Context Protocol (MCP) server (list_notebooks, search_notebook_chunks, get_notebook_transcript) supporting both Dynamic OAuth 2.1 (IndieAuth) for interactive desktop agents (Claude Desktop, Cursor) and revocable Static Bearer API Keys (sn_...) for headless scripts and CLI automation.
Filename & Page-Style Driven Custom AI Prompts
Drop custom Markdown prompt templates into your prompts_dir (ocr/daily/prompt.md, summary/meeting_notes/prompt.md, or ocr/style_cornell/prompt.md), and the server automatically routes matching notebooks (by filename stem or page stationery style) through your custom instructions while injecting native device page keywords (PAGE_KEYWORDS). See AI Pipeline, Custom Prompts & MCP Server for full prompt routing details.
Live Calendar Feeds (.ics), Webhooks & Prometheus/Grafana Observability
- iCalendar (
VTODO) Feeds: Tasks created on your Supernote tablet are automatically exposed as an RFC 5545.icsfeed (/api/schedule/feed.ics) for Apple Reminders/Calendar, Google Calendar, Outlook, and Home Assistant. - Signed Outbound Webhooks: Receive HMAC-SHA256 signed HTTP callbacks (
note.sync_completed) the instant a notebook finishes syncing. - Production Monitoring: Built-in
/metricsendpoint with reference Prometheus recording/alerting rules and a 13-panel Grafana dashboard.
Offline Binary .note Parser, Converter & Cooperative Markdown Authoring
You don't need to run a server to use supernote. Use the CLI (supernote analyze, supernote convert, supernote create, supernote merge, supernote reconstruct) or supernote.notebook to export .note files to PDF/PNG/SVG, or programmatically generate and append stylized stationery templates, simulated handwriting, and hidden flashcard layers (LAYER1–LAYER3) onto .note files.
Documentation Map
| Guide / Reference | What You'll Learn |
|---|---|
| Self-Hosting & Quickstart Guide | Deploying via Python or Docker Compose, bootstrapping your Admin account, pairing your Supernote tablet, reverse proxies (Caddy/Nginx), and SQLite/blob backups. |
| Configuration Reference | Complete reference for config.yaml and SUPERNOTE_* environment variables, storage quotas, Gemini, named local OpenAI-compatible servers (Ollama/vLLM), and macOS Vision. |
| AI Pipeline, Custom Prompts & MCP Server | How the 5-stage OCR/Embedding/Summary pipeline works, routing custom prompts by filename or page_style, static API keys (sn_...), and connecting Claude or Cursor via MCP. |
| OCR Provider Benchmarks | 30-page Supernote handwriting & Bullet Journal benchmark comparing Gemini 3.5 Flash-Lite, Gemini 3.6/3.8 Flash, Supernote On-Device (MyScript iink), and Apple Vision. |
| Prometheus & Grafana Monitoring | Scraping /metrics, importing the reference Grafana dashboard, and configuring Prometheus recording and alerting rules. |
Webhooks, iCal (.ics) & Integrations |
Subscribing to RFC 5545 VTODO task feeds in Apple/Google/Outlook calendars and Home Assistant, plus HMAC-signed outbound webhooks. |
CLI Reference (supernote) |
Every command and flag for local .note conversion and authoring (create), server hosting (serve), cloud operations, API keys (cloud keys), search, and admin management. |
| Server & Sync Protocol Architecture | Deep dive into the Supernote Private Cloud wire protocol (challenge-response auth, 3-step CAS sync, Socket.IO) and our Python/SQLite implementation. |
Binary .note Format Specification |
Byte-level layout of Supernote .note and .mark headers, footers, metadata blocks (schema.py), layers, and RATTA_RLE / RATTA_RLE_X2 bitmap compression. |
Python Package & API Reference
supernote.server: Open Private Cloud Server, Services & Routes (supernote.server.config.ServerConfig,supernote.server.app.create_app)supernote.client: Async Cloud Client SDK (supernote.client.Supernote,supernote.client.device.DeviceClient,supernote.client.schedule.ScheduleClient,supernote.client.extended.ExtendedClient)supernote.notebook: Notebook Parser, Converter & Authoring Library (supernote.notebook.load_notebook,supernote.notebook.PdfConverter,supernote.notebook.author.NotebookAuthor,supernote.notebook.renderer.MarkdownRenderer)supernote.models: Protocol DTOs & VOs (supernote.models.file_device,supernote.models.schedule,supernote.models.summary,supernote.models.extended,supernote.models.webhook)
Choosing Your Private Cloud Experience
While the official Supernote Private Cloud by Ratta provides a production-grade managed sync experience, Open Private Cloud offers a lightweight self-hosted alternative with opt-in AI enhancement:
| Capability | Official Private Cloud (Ratta) | Open Private Cloud (Bare Sync Mode) | Open Private Cloud (Full Knowledge Hub) |
|---|---|---|---|
| Core Sync | Robust & Validated | 100% OpenAPI Compliant | 100% OpenAPI Compliant |
| Memory Footprint | High (~2 GB+ RAM required) | Low (~135 MB flat RAM) | Low (~160 MB idle / ~310 MB active peak) |
| Stack & Database | Multi-container (Java + Redis + MariaDB) | Python / asyncio + single SQLite file |
Python / asyncio + single SQLite file |
| Web UI & Tasks | Built-in Web UI | Disabled (web_enabled: false) |
Editorial Web UI + .ics Calendar Feeds |
| AI Analysis | Basic OCR (Device-side) | None | Optional: AI Transcriptions & Synthesis (Gemini or Local Models) |
| Search | Path / Filename | Path / Filename | Optional: Cross-Notebook Semantic Concept Search |
| AI Agent Access (MCP) | None | Disabled (mcp_enabled: false) |
Built-in OAuth 2.1 & Static API Key MCP Server |
Quick Start
For full deployment instructions (including docker-compose.yml, Caddy/Nginx reverse proxies, and backups), see the Self-Hosting & Quickstart Guide.
Launch the Server
Option A: Lite Sync Server (Zero-Config)
No API keys or external services required. Runs locally with SQLite (~160 MB idle, ~310 MB peak when rendering notebook previews).
- Using Python:
pip install "supernote[server]" supernote serve
- Using Docker:
docker build -t supernote . docker run -d \ -p 8080:8080 \ -v $(pwd)/storage:/data \ --name supernote-server \ supernote
- Bare Sync Mode (~135 MB flat RAM): To run a minimal headless sync server without the Web UI, background PNG rasterization, or MCP server, disable
web_enabledandmcp_enabledvia environment variables orconfig.yaml:SUPERNOTE_MCP_ENABLED=false SUPERNOTE_WEB_ENABLED=false supernote serve
# config/config.yaml mcp_enabled: false web_enabled: false
Option B: AI & Knowledge Hub (With Gemini, Recommended)
Enables handwriting transcription, summarization, and semantic search using gemini-3.5-flash-lite by default. Requires a Google Gemini API Key (available on both a $0 Free tier that trains on API data and a Paid tier that does not; see Choosing an AI Provider).
- Using Python:
export SUPERNOTE_GEMINI_API_KEY="your-gemini-api-key" pip install "supernote[all]" supernote serve
- Using Docker:
docker build -t supernote . docker run -d \ -p 8080:8080 \ -v $(pwd)/storage:/data \ -e SUPERNOTE_GEMINI_API_KEY="your-gemini-api-key" \ --name supernote-server \ supernote
Option C: AI & Knowledge Hub (With a Local Model)
Runs the same features against a model server on your own hardware, such as Ollama, llama.cpp, vLLM, or LM Studio, so your notes never leave your network. You need a vision-capable chat model for OCR and summaries, and an embedding model for search. Configure the openai provider in config/config.yaml (/data/config/config.yaml in Docker) as described in the Configuration Reference. On macOS, Apple's native Vision framework (provider: apple with vision-accurate) is also supported as an offline fallback for basic prose search, though it is not recommended for structured notes, diagrams, or Bullet Journal Rapid Logging (see Choosing an AI Provider).
When the server runs in Docker and Ollama runs on the host, use http://host.docker.internal:11434/v1 as the base_url (on Linux, also pass --add-host=host.docker.internal:host-gateway).
Bootstrap Your User
Once the server is running, register your administrator account (the first user registered automatically receives Admin privileges and closes public registration):
- Using Python CLI:
# Create the initial admin account supernote admin --url http://localhost:8080 user add you@example.com # Authenticate your CLI supernote cloud login you@example.com --url http://localhost:8080
- Using Docker CLI:
docker exec -it supernote-server supernote admin --url http://localhost:8080 user add you@example.com
Connect Your Device
- On your Supernote, go to Settings > Sync > Supernote Cloud and select Private Cloud.
- Enter your server URL (e.g.,
http://192.168.1.5:8080). - Log in with the email and password you created in Step 2.
- Tap Sync to begin syncing your notes.
Explore Your Insights
Once your notes sync and process, you can view the AI synthesis from the terminal, browser, or MCP server (http://<your-ip>:8081/mcp):
# Get a high-level summary and transcription
supernote cloud insights /NOTE/Note/Journal.note
# Semantic search across all notebooks
supernote cloud search "What were my project goals for February?"
Choosing an AI Provider
We recommend Google Gemini (gemini-3.5-flash-lite by default, or gemini-3.6-flash) for OCR, summarization, and embeddings based on our 30-page Supernote OCR benchmark and public multimodal evaluations. Quality for summaries and handwriting OCR extraction may be lower when using smaller local or other OpenAI-compatible models.
Supernote OCR Benchmark (evals/ocr/)
Results across 30 .note pages (25 real handwritten pages covering Bullet Journal Rapid Logging, two-column diagrams and guitar tablature, French/Spanish/Turkish notes, and overlapping strokes, plus 5 synthetic Bullet Journal fixtures; full breakdown in evals/ocr/RESULTS.md):
| Provider / Model | Pages | Word F1 | Mean CER | Content CER | Handwritten BuJo Symbols | All BuJo Symbols | Median Latency |
|---|---|---|---|---|---|---|---|
Gemini 3.5 Flash-Lite |
30 | 99.0% | 2.3% | 1.4% | 85.5% | 95.8% | 1.18 s |
Gemini 3.6 Flash |
30 | 99.7% | 4.2% | 3.6% | 90.5% | 97.3% | 6.63 s |
Gemini 3.8 Flash |
30 | 96.5% | 5.6% | 5.4% | 90.5% | 97.3% | 9.93 s |
Supernote On-Device (MyScript iink 4.1.0) |
25 | 75.5% | 17.2% | 16.7% | 18.6% | 18.6% | 0.91 s |
Apple Vision Accurate |
30 | 73.7% | 21.7% | 20.7% | 4.5% | 49.7% | 0.41 s |
Apple Vision Fast |
30 | 17.8% | 72.8% | 70.7% | 0.0% | 8.7% | 0.13 s |
- Gemini:
gemini-3.5-flash-lite(the default) reads nearly every word accurately (99.0% Word F1, 0.0% Median CER) in ~1.2 s per page for under$0.001per page, while preserving Bullet Journal signifiers (•,X,>,<,o,-,*), nested indentation (100%), and~~strikethroughs~~(100%).gemini-3.6-flashreaches 99.7% Word F1 and 2.3% multilingual CER; its higher diagram CER comes from adding ASCII|fret dividers on guitar tablature (99.3% Word F1 on those pages). - Apple Vision: Supported on macOS (
vision-accurate) as an offline fallback for plain English prose (93.0% Word F1 onenglish_prose), not as a recommended provider. Overall accuracy matches the tablet's built-in recognizer (73.7% vs 75.5% Word F1), multi-column layouts are read out of order (30.2% diagram CER), and because Vision cannot take a prompt it drops almost all handwritten Bullet Journal task symbols (4.5% recall) and strikethroughs (0%). Avoidvision-fast, which only detects machine-printed text.
Multimodal Benchmarks: Gemini vs. Claude
Recent results (Jun–Nov 2026) from Artificial Analysis (MMMU-Pro), Vals AI (MMMU-Pro and MortgageTax scanned-form extraction), and LMArena Vision across model tiers:
| Model | Release | API Price (In / Out per 1M) | LMArena Vision (Elo) | MMMU-Pro (AA) | MMMU-Pro (Vals) | MortgageTax (Vals) |
|---|---|---|---|---|---|---|
Gemini 3.5 Flash-Lite |
Jul 2026 | $0.30 / $2.50 |
1268 | 79.0% | 83.6% | 68.7% |
Claude Haiku 5.5 |
Oct 2026 | $0.10 / $0.50 (≤100k) |
— | — | — | — |
Gemini 3.6 Flash |
Jul 2026 | $0.75 / $3.75* |
1297 | 83.2% | 88.4% | 67.6% |
Gemini 3.8 Flash |
Sep 2026 | $0.75 / $3.75* |
1316 | 85.6% | 89.1% | 65.3% |
Claude Sonnet 5 |
Jun 2026 | $2.00 / $10.00 |
1272 | 77.3% | 83.0% | 70.0% |
Claude Sonnet 5.5 |
Sep 2026 | $2.00 / $10.00 |
1283 | — | — | — |
Gemini 4 Carbon |
Oct 2026 | $4.00 / $20.00 |
— | 89.3% | — | — |
Claude Opus 5 |
Jul 2026 | $5.00 / $25.00 |
1317 | 84.7% | 89.9% | 72.1% |
Claude Opus 5.5 |
Sep 2026 | $4.00 / $20.00 |
1324 | 87.7% | — | — |
*Promotional rate through Dec 31, 2026 ($1.50 / $7.50 standard; 50% lower via Batch API).
Gemini 3.5 Flash-Lite matches Claude Sonnet 5 on MMMU-Pro at roughly one-sixth the price, while Gemini 3.6 Flash and Gemini 3.8 Flash score ahead of Claude Sonnet 5 and 5.5 on LMArena Vision and close to Claude Opus 5 at a fraction of the cost.
API Privacy, Training & Data Retention
Google Gemini offers both a Free tier that trains on API data and a Paid tier (enabled by linking a Cloud Billing account to your Google AI Studio project) that does not. Commercial API traffic on OpenAI and Anthropic is excluded from training by default:
| Provider & Tier | Used to Train Models? | Default API Log Retention | Policy Citations |
|---|---|---|---|
| Gemini API (Paid tier) | No (Google acts as Data Processor) | Abuse monitoring; optional AI Studio project logs default to 55 days (configurable to 7 days or off; ZDR on Enterprise/Vertex AI) | Paid Services Terms, Logs Policy, Google DPA |
| Gemini API (Free tier) | Yes (outside EEA, UK, and Switzerland; includes human review) | Retained for product improvement (avoid for private or sensitive notes) | Unpaid Services Terms |
OpenAI API (api.openai.com) |
No (unless opted in) | Up to 30 days for abuse monitoring; ZDR available for eligible orgs | OpenAI Data Controls |
Anthropic Claude API (api.anthropic.com) |
No under Commercial Terms | Not retained by default on standard endpoints, or up to 30 days on covered models; ZDR available | Commercial Terms, Training Policy, Data Retention |
Local models (Ollama, llama.cpp, vLLM, apple) |
No | Local only; notes never leave your server | — |
Installation
# Install specific components
pip install supernote # Standalone .note parser, converter & authoring only
pip install "supernote[server]" # + Open Private Cloud server, Web UI, AI & MCP
pip install "supernote[client]" # + Async Cloud API client & cloud/admin CLI
pip install "supernote[server,apple]" # + macOS Vision OCR fallback
# Full installation (recommended for server users)
pip install "supernote[all]"
CLI Usage
See the CLI Reference for all commands and flags.
# Server & Admin
supernote serve # Start Open Private Cloud
supernote admin user list # Manage users
supernote admin user set-quota you@example.com 5G # Set per-user storage quota
# Offline Notebook Parsing, Conversion & Authoring
supernote convert journal.note journal.pdf -a # Export .note to PDF
supernote create daily.note --markdown "# Daily Log\n\n* - [ ] Ship release"
# Cloud File Operations & AI Synthesis
supernote cloud ls / # List remote files
supernote cloud download /NOTE/Note/Journal.note # Download to local machine
supernote cloud insights /NOTE/Note/Journal.note # View AI summary & OCR transcript
supernote cloud search "Q1 roadmap" # Semantic vector search
# Static API Keys & Headless Automation
supernote cloud keys create "Automation" --expires-days 90
supernote cloud keys list
supernote cloud keys revoke <key-id>
supernote cloud --url http://localhost:8080 --api-key sn_... ls /
Headless scripts can also set SUPERNOTE_URL and SUPERNOTE_API_KEY in the environment instead of running supernote cloud login.
Notebook Operations & Developer API
Parse, Convert & Author .note Files (supernote.notebook)
See the Notebook Parser, Converter & Authoring Guide for full details:
from pathlib import Path
from supernote.notebook import PdfConverter, load_notebook
from supernote.notebook.author import create_notebook
# Convert an existing .note file to a multi-page PDF with clickable links
note = load_notebook("journal.note")
pdf_bytes = PdfConverter(note).convert(-1, enable_link=True)
Path("journal.pdf").write_bytes(pdf_bytes)
# Author a new .note file directly from Bullet Journal Markdown
note_bytes = create_notebook(
"# Daily Log\n\n* - [ ] Review ==architecture spec== #engineering",
style="stylized",
background="dots",
auto_toc=True,
)
Path("daily.note").write_bytes(note_bytes)
Automate Cloud Sync, Tasks & Search (supernote.client)
See the Python Client SDK Guide for the complete async API:
from supernote.client import Supernote
from supernote.client.extended import ExtendedClient
# Authenticate with email/password (`Supernote.login`) or a static API key (`sn_...`)
async with Supernote.from_token("sn_...", host="http://localhost:8080") as sn:
files = await sn.device.list_folder("/")
ext = ExtendedClient(sn.client)
hits = await ext.search("architecture trade-offs", top_n=5)
for hit in hits.results:
print(f"{hit.file_name} (p.{hit.page_index + 1}): {hit.score:.0%}")
The notebook parser originally started as a fork of supernote-tool with lighter dependencies and modern type hints. All credit goes to the original authors for providing the foundational parser utility.
Contributing & Community
We welcome issues, feature requests, and pull requests on GitHub (allenporter/supernote)! Please see our Contributing Guide and Database Migration Guide for details on local development (./script/bootstrap, ./script/test, ./script/lint, ./script/docs), Ephemeral Mode (supernote serve --ephemeral), and AI Skills.
Community Projects
- jya-dev/supernote-tool — Original parser foundation.
- awesome-supernote — Curated resource list.
- sn2md — Supernote to text/image converter.
Metadata
Release files for supernote 0.26.0
For a detailed explanation of source distributions (sdists) and built distributions (wheels), please see the package formats documentation.
Source distribution (sdist)
| File | Size | Uploaded | Provenance | Actions |
|---|---|---|---|---|
| supernote-0.26.0.tar.gz | 956.2 kB | With provenance | Details for supernote-0.26.0.tar.gz |
Built distribution (wheel)
| File | Interpreter | ABI | Platform | Actions Reset |
|---|---|---|---|---|
| supernote-0.26.0-py3-none-any.whl | Python 3 | none | any | Details for supernote-0.26.0-py3-none-any.whl |
Total release size: 2.0 MB
Release files / supernote-0.26.0.tar.gz
| Download URL | supernote-0.26.0.tar.gz |
|---|---|
| Size | 956.2 kB |
| Tags | Source |
|
SHA-256 checksum How to use checksums |
aa6944711192eb26f4012e6b1f7ac3e1b080072e7b128f8c05a1463fd22604f9
|
|
BLAKE2b-256 checksum How to use checksums |
4b16cea9d2e5e01306962c5305862853f1833cca01848e98e5bde9f543e12cf1
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 16, 2026.
Transparency logRelease files / supernote-0.26.0-py3-none-any.whl
| Download URL | supernote-0.26.0-py3-none-any.whl |
|---|---|
| Size | 1.1 MB |
| Tags | Python 3 |
|
SHA-256 checksum How to use checksums |
0972cfd06995713c0f41ae8b3b4cff61a6ec11aa21a5c4e327983d8868850131
|
|
BLAKE2b-256 checksum How to use checksums |
ecee6bc1ed21a97934cd9bb5e669725869d2221f6abe6c14561c5fcf207621b7
|
| Upload date | |
|
Uploaded using Trusted Publishing? What is trusted publishing? |
Yes |
| Uploaded via |
twine/7.0.0 CPython/3.13.16
|
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 16, 2026.
Transparency logSecurity for supernote 0.26.0
Provenance
Provenance describes where a file came from. On PyPI, provenance is shared via attestations, which provide a verifiable record of the build or publishing details. View details, limitations and caveats.
Full provenance
Both files in this release have been uploaded with provenance metadata.
Attestation for supernote-0.26.0.tar.gz
PyPI Publish Attestation
PyPI verified that this artifact, at this checksum, originated from the publisher listed below.
Signed by GitHub Actions, verified by PyPI on Nov 16, 2026.
Transparency logA note on package security
PyPI is a repository for distributing packages, not a service for auditing them. While third-party security researchers and our staff do their best to detect and remove malicious packages, PyPI cannot guarantee the quality, security or suitability of any given package.
To learn more about protecting yourself from malicious actors, visit our package security documentation.