EdgeProc

A Python library for sending data files to many devices: each device checks they came from you, downloads only what changed, and can search them offline.

Try it in a minute: pip install "edge-proc[bundles]", then run the commands under Try it. No server or account needed.

Say your app needs the same data on many devices you don't control: a product catalog, a search index, a price list, a small AI model. The usual way is to put the file on a web server and have each device download it. That has two problems. A device can't tell if the file was corrupted or swapped on the way, so it may quietly give wrong answers. And when one line changes, every device downloads the whole file again.

EdgeProc handles both. You publish a folder once, on your own machine: it cuts the files into small pieces and signs a list of them with your private key. You upload the result to any plain web server. Each device downloads only the pieces it doesn't have yet, checks them all against your public key, and refuses the whole update if anything is wrong. Once the data is on the device, EdgeProc can also search it there, with no network.

Technical docs: Architecture · Getting started for developers · Full walkthrough with search · Configuration · Operations and security

Try it

This uses local folders instead of a web server, so you can see every step. You need Python 3.13 or newer.

1. Install EdgeProc from PyPI in a fresh virtual environment. The [bundles] extra adds signing and syncing; it is small.

   python3.13 -m venv .venv && source .venv/bin/activate
   pip install "edge-proc[bundles]"
   

2. Make a two-product catalog, a key pair, and publish it. Then pull it onto a "device" folder.

   mkdir -p demo/src && cd demo
   echo '{"p1": "red running shoes", "p2": "waterproof hiking boots"}' > src/catalog.json
   edgeproc keygen --out keys
   edgeproc publish --src src --origin-dir origin --key keys/private.key --bundle-id catalog --version 1.0.0 --pretty
   edgeproc sync --base-url origin --cache-dir cache --key keys/public.key --materialize-to device --pretty
   cat device/catalog.json
   

   wrote keys/private.key and keys/public.key
   key_id 87207cf94bb9c28b
   published v1.0.0 manifest=71d733a8dddb
   synced v1.0.0 manifest=71d733a8dddb chunks_fetched=1 chunks_reused=0 bytes_fetched=70
   {"p1": "red running shoes", "p2": "waterproof hiking boots"}
   

   keys/private.key stays with you. keys/public.key is what you give every device. origin/ is what you would upload to a web server. Your key_id will differ; the rest repeats exactly.

3. Add one file and publish a new version. The device downloads only the new piece.

   echo 'spring price list' > src/prices.txt
   edgeproc publish --src src --origin-dir origin --key keys/private.key --bundle-id catalog --version 1.0.1 --pretty
   edgeproc sync --base-url origin --cache-dir cache --key keys/public.key --materialize-to device --pretty
   

   published v1.0.1 manifest=18499ab39ead
   synced v1.0.1 manifest=18499ab39ead chunks_fetched=1 chunks_reused=1 bytes_fetched=27
   

4. Try to sync without the public key. EdgeProc refuses; there is no "trust it this once" mode.

   edgeproc sync --base-url origin --cache-dir other-device --pretty; echo "exit=$?"
   

   [config.missing] no trust root: pass --key or set EDGEPROC_TRUST_ROOT_PUBKEY_PATH (refusing to sync)
   exit=1
   

5. Corrupt one piece on the "server" and sync onto a new device. The whole update is refused.

   printf 'tampered' > "origin/chunk/$(ls origin/chunk | head -1)"
   edgeproc sync --base-url origin --cache-dir new-device --key keys/public.key --pretty; echo "exit=$?"
   ls cache new-device
   

   [bundle.integrity_failed] sync failed: stored chunk failed to decompress
   exit=1
   cache:
   active
   chunks
   manifests
   
   new-device:
   chunks
   manifests
   

This is real output from edge-proc 0.5.1, installed from PyPI on 2026-09-25.

To see on-device search as well, clone the repo and run bash examples/run_loop.sh. It builds a search index over a 15-product catalog, ships it together with the search model, and searches on the "device" with no network. The query "trail running shoes for the morning" returns the "Trailhead Pro running shoes" first. The first run downloads an 87 MB search model and about 1 GB of dependencies (PyTorch and FAISS); it took about 2 minutes. The full walkthrough does the same thing one step at a time.

How it works

edgeproc publish cuts each file into pieces of about 64 KB. Each piece is named by a fingerprint of its contents (a SHA-256 hash), so a changed or damaged piece no longer matches its name. A list of all the pieces is fingerprinted too, and one small file that points to that list is signed with your private key. edgeproc sync checks that signature against the public key the device was given, downloads only the pieces it is missing, and checks each one. Only when every check passes does the new version go live; otherwise the device keeps the last good version.

If you publish a SQLite database (.sqlite, .sqlite3, or .db), edgeproc publish first checks it, compacts it onto 4 KB pages, and zeroes the header counters that SQLite changes on every write. That way, changing one row in a 10 MB SQLite catalog changes only the affected piece instead of shifting the whole file.

For search, EdgeProc turns text into vectors with a small model and looks them up in a FAISS index, optionally mixed with keyword search. Ship the model inside the same signed release and the device never needs to download it. Without a local model, search refuses to run rather than fetch one.

How it fits with the related projects

• edgeproc-core (Python, on PyPI) is a smaller library underneath this one. EdgeProc uses two things from it: a common interface for vector search indexes, which EdgeProc's FAISS index implements, and the stable error codes, like bundle.integrity_failed. It installs automatically with EdgeProc. You would only use it on its own to keep each customer's results apart in a shared search index.
• @gainratio/browser (TypeScript, for web pages; previously @edgeproc/browser) does the checking half of EdgeProc inside a browser tab. It reads the same signed format that edgeproc publish writes, so you can publish a folder or a .sqlite database with this library and load it in a web page (for example, an offline product catalog or a local reference database).
• privacy-core (published on npm as @gainratio/privacy-core; previously @edgeproc/privacy-core) is a separate project despite the old shared scope. It hides card numbers and similar IDs from AI prompts in the browser, and does not use EdgeProc.

What it does not do

• It is not a service. There is nothing to sign up for. It is a library and a command-line tool that your app and your build machine run.
• Data flows one way. One publisher signs; many devices read. It is not a two-way folder sync like Dropbox.
• It can't protect a device or build machine that is already compromised, or a private key that was stolen. You can revoke a key, but each device only learns about it when you update its key ring. There is no online revocation list.
• The memory budget is not a hard memory limit. EdgeProc refuses new work when the declared budget is used up, but it does not cap what FAISS or PyTorch actually allocate.
• Search is heavy to install. The [localvec] extra pulls in PyTorch and FAISS, about 950 MB.
• CI tests on Linux only. It is used on macOS too, but Windows and macOS are not tested in CI.
• Search quality on your data is not measured. The default model is a small general one (all-MiniLM-L6-v2). Check it against your own data.

When to use something else

If you need	Use
No code on the device, always-fresh data, and per-search cost is fine	A hosted search or vector database API
One small file that rarely changes, from a server you fully trust	A plain HTTPS download with a checksum
Several signing roles, threshold signing, or online revocation	A full software-update framework such as TUF
Two-way sync between machines you control	rsync or a file-sync tool
The same data on devices you don't control, checked before use, with small updates and offline search	EdgeProc

Install

EdgeProc is on PyPI as edge-proc. You import it as edgeproc, and the command is also edgeproc. The heavy parts are optional extras:

pip install edge-proc                    # the command-line tool and the task router
pip install "edge-proc[bundles]"         # + publish and sync (signing, checking, small updates)
pip install "edge-proc[localvec]"        # + on-device search (FAISS and PyTorch, about 950 MB)
pip install "edge-proc[localvec,bundles]"

This README documents EdgeProc 0.5.1. It needs Python 3.13 or newer and installs edgeproc-core>=0.4.3 from PyPI. Release notes are in the CHANGELOG.

Develop

git clone https://github.com/gainratio/edge-proc.git
cd edge-proc
uv sync --all-extras
uv run poe gate

The last command runs lint, type checks, the complexity limit, and the full test suite. It took about 3.5 minutes. Getting started for developers covers setup, a map of the code, a first change, and how to open a pull request.

More detail

• Getting started for developers: from a fresh clone to a green local build and your first change.
• Full walkthrough: build a search index, ship it with its model, and search it on the device, step by step. Also shows the Python API.
• Architecture: the modules, the bundle format, the security model, key rotation, and what the tests prove.
• Explore the interactive architecture map: the same, as a clickable diagram that works offline.
• Configuration: every setting and environment variable.
• Operations: threat model, privacy and data flow, recovery, the key rotation runbook, measured performance, and how releases are made.
• Roadmap: what is planned and not built yet.
• CHANGELOG: what changed in each release.
• CONTRIBUTING and SECURITY: how to contribute and how to report a vulnerability privately.

License

MIT. See LICENSE. To cite EdgeProc, use CITATION.cff.

EdgeProc (also written edge-proc and edgeproc) lives at gainratio/edge-proc, with a project page at edge-reco.com/edgeproc. It is not affiliated with any other product or company named "EdgeProc".