Self-hosted catalogue of your skills, agents and MCP servers, searchable by meaning over MCP
Runs on your machine from its package; the steps are below.

At most: only reads or looks up information.
Odel charges nothing per call. The publisher may charge for its own service.
Runs locally over stdio
This server isn't hosted — your MCP client launches it from a package registry. Use one of the commands below, or drop the config into your client (e.g. Claude Desktop).
Run
docker run -i --rm ghcr.io/bobpanil/vivatlas:1.0.13
MCP client config
{
"mcpServers": {
"vivatlas": {
"command": "docker",
"args": [
"run",
"-i",
"--rm",
"ghcr.io/bobpanil/vivatlas:1.0.13"
]
}
}
}
You collect tools. Skills for Claude, agents, MCP servers, a scraper you starred at midnight, a design kit somebody linked on Reddit. They end up scattered across GitHub stars, a self-hosted Gitea, browser bookmarks and a notes app — and three months later you remember there was a thing for that and can't find it.
VIVATLAS is the shelf. Point it at your repositories and the links you clip, and it turns each one into a card: a name, a description written for a person rather than a compiler, a picture, tags. Then it puts them all in one place you can search by meaning, not just by name. Ask it for "something that audits a site for accessibility" and it hands you three candidates and tells you why.
It runs on your own machine — a NAS, a spare box, a laptop — it's built for more than one person, and it only ever reads. Public repositories in, nothing written back to Git, ever.
Try it before you install it. demo1.bobpanil.com is a public VIVATLAS. Press Try it on the sign-in page for a private account of your own, then browse the catalogue, search it, and add links or notes. The account and everything you add are deleted after 24 hours. The demo runs without AI, so the cards you add are plain and search goes by words, not by meaning. It always runs the newest release.
Nobody browses a wall of grey rectangles. VIVATLAS finds each project's own picture — the banner at the top of its README, a logo in the repo, or the card GitHub draws for it — and throws out the things that only look like pictures: build badges, sponsor logos, "buy me a coffee" buttons, the owner's avatar. When a project has several real candidates, the same AI that writes the description looks at them side by side and picks the one that actually shows the thing.
And when a project has nothing at all, VIVATLAS draws a cover itself — the name set in the house typeface on a colour and a pattern that come from the name, so every cover is different and they all belong together. No model, no cost, never grey.
Open one and you get the same project three ways: a one-liner, a paragraph for a person, and a technical note for when you're deciding whether to use it. Tags with their confidence, what it's for, where it came from and when it last changed — and the buttons to rescan it, file it, or copy the link. The descriptions are written by whichever AI you've configured, from the repository's own README.
Two indexes work together: full-text for words, vectors for meaning. Type in Russian and find an English tool. Describe a task instead of naming a tool, and the recommender returns three options with a sentence each on why — or an honest "nothing here fits", which is rarer in software than it should be.
Browser extension (Chrome, Edge, Brave, Firefox): the page you're on, or a pasted link, into your catalogue in one click, public or private. → extension/README.md
Android app: Share → VIVATLAS from any app. A small sheet shows what you're saving and lets you choose private or public before it goes. Signs in by scanning a QR from a browser that's already signed in, so a long password never meets a phone keyboard. → android/README.md
The web form: a link, a site, a screenshot or a reel → candidates with stars → pick → a card in your catalogue.
Installing the Android app: from F-Droid, which keeps it updated. Or grab
vivatlas-android-<version>.apk from the latest release, or add this repository to Obtainium and it updates itself. Releases are signed with one key for the life of the app, so each installs over the last.
Each person has their own sign-in, their own private cards and personal folders. The shared catalogue is common to all and curated by whoever runs the place. A card can move between the two — it's the owner's call — and you can keep favourites, watch a feed of what changed, and see which cards have gone stale upstream.
VIVATLAS is an MCP server. Connect it to ChatGPT, Claude, or anything else that speaks MCP, and "which of my tools does X?" gets answered from your catalogue. There's a plain REST API too. → docs/MCP.md
Three languages (English, Russian, Hebrew — right-to-left done properly). Light, dark, OLED and follow-the-system themes. Two-step sign-in with backup codes. Invitations by link or email, or open registration if you prefer. Gitea and GitHub as sources, rescanned daily. An admin panel for people, folders, email and AI keys — all changeable without a restart. And upstream / update, which compare a card with where it came from and bring in a new version only where you haven't touched the file.
The quickest way is the container:
docker run -d --name vivatlas -p 8710:8710 -v vivatlas_data:/data ghcr.io/bobpanil/vivatlas:latest
Open http://localhost:8710, and the first person through /setup becomes the owner. Then Admin → Sources: give it a GitHub account or a Gitea address, press Scan now, and watch the cards arrive.
Running from source instead:
python -m venv .venv
.venv/bin/python -m pip install -e ".[dev]" # Windows: .venv\Scripts\python.exe
cp .env.example .env # set SECRET_KEY; everything else is optional
.venv/bin/python -m vivatlas.cli init-db # create or update the database
.venv/bin/python -m vivatlas.cli scan # fetch repositories
.venv/bin/python -m vivatlas.cli serve --host 0.0.0.0 --port 8710
--host 0.0.0.0 lets a phone on the same Wi-Fi reach it. After updating the code, run init-db again — serve doesn't migrate on its own (the container does, on every start).
For a proper home, there's a walkthrough for TrueNAS SCALE in docs/DEPLOY-TRUENAS.md — it also covers the two settings that matter once VIVATLAS sits behind a proxy or a tunnel.
Everything optional lives in .env (or the container's environment); .env.example annotates the lot. The ones people ask about:
| Setting | What it's for |
|---|---|
GOOGLE_API_KEY, OLLAMA_URL | The AI that writes descriptions and picks pictures. Either works; both is fine. Also settable in the admin panel. |
TRUSTED_PROXIES | If Cloudflare, Traefik or nginx terminates TLS in front of VIVATLAS. Without it, cookies go out without Secure. |
IMAGE_MODEL | Have pictures drawn for cards that offer none — a Google image model (needs billing) or pollinations:flux (free, keyless, soft). Empty means the designed covers, which is the default and looks better than you'd think. |
VIVATLAS is a viewer, not an owner. It reads public Git repositories and helps you find and view the tools inside them. It does not own — and claims no rights in — any of the repositories, code, names or trademarks it catalogues or links to. Those belong to their authors and are governed by their own licences.
It writes nothing to Git and does not scan private repositories. That's a rule, not a setting: there's no toggle for it anywhere.
Licence: the server — everything under src/ — is Business Source License 1.1: free to use, deploy and modify, including commercially; you may not sell or resell VIVATLAS itself as a product or a hosted service. It converts to Apache-2.0 on 2030-07-21. The Android app (android/) is Apache-2.0, so it can be built and carried by F-Droid. The browser extension (extension/) is Apache-2.0 too.
.venv/bin/python -m pytest
.venv/bin/python -m ruff check src tests
src/vivatlas/
config.py settings from .env
runtime_settings.py overrides from the database on top of .env (edited from the admin panel)
db.py connection to SQLite
models.py tables
migrate.py init-db: add missing columns/indexes, rebuild search
security.py passwords, secret encryption, secret key
twofactor.py two-step sign-in (TOTP + backup codes)
auth.py, auth_web.py sign-in, registration, invitations, password reset
qrlogin.py sign a phone in by a one-time QR shown on a signed-in browser
admin_web.py admin panel (people, access, email, integrations)
settings_web.py personal settings, avatars, sources, folders
web.py catalogue, cards, adding
api.py app assembly, REST, background loops, static
filters.py visibility: own + shared
categories.py folder permissions (shared/personal)
i18n.py, translations*.py translations (en/ru/he), RTL
mailer.py emails (password reset, invitations)
avatars.py uploaded photo → square WebP
previews.py the picture on a card: find it, fetch it, fit it — or draw a cover
ai/ the models: google.py (text, vision, images), ollama.py, pollinations.py
usericons.py default avatar set (static/usericons)
scanner.py, indexer.py scanning + the private-repo rule, index
mcp_server.py MCP server for AI assistants
cli.py terminal commands (init-db, scan, serve, previews, embed, upstream…)
providers/
base.py common interface to a Git host (the "socket")
gitea.py Gitea
github.py GitHub (an account's public repositories)
ext_api.py JSON API for the browser extension and the Android app (/api/ext)
templates/, static/ pages and styles (hand-written CSS, no build step)
extension/ browser extension for Firefox and Chromium (Manifest V3)
android/ Kotlin WebView shell: share target, QR sign-in
Adding another Git host: implement the providers/base.py interface in a new provider and wire it in providers/__init__.py. The rest of the code stays the same.