The web is our common land.
Cytir is Welsh for common land: ground that belongs to everyone who works it. We are building a web search index that way. It is made for AI agents. Download it, run it on your own ground, extend it with your own sources. What you search stays with you.
Why
Web search was built for people. Agents need their own.
For decades web search has belonged to a handful of companies indexing hundreds of billions of pages. An open alternative at that scale was never realistic: the infrastructure cost is enormous, and the index would be far too large to share. So we asked what an agent actually needs.
Precision first
An agent answering a legal question has to read the exact statutory text before it answers. It needs the authoritative source itself, not a page about it.
Fresh, not months old
A model is trained on data that is months old by the time it ships. Search is how an agent sees what changed since.
Sovereign all the way down
Open-weight models make sovereign LLMs possible. An agent that still sends every query outside your perimeter is only half sovereign.
Yours to keep and to change
The code and the index are open. A foundation keeps the core index honest, and anyone can run it, fork it or add their own sources.
How it fits
Agents don't need the web the way people do.
Two differences let us redesign the index from scratch, for size and for sharing.
The model already carries a compressed web
So the index only has to fill in what the model lacks: the exact text, the current state, the authoritative URL.
A good entry point is enough
Agents don't need the perfect page in the top three results. Given a good entry point, they walk the site themselves and find the best page in seconds.
Sizes
Three sizes of common.
The same software runs every size. The index lives on NVMe; RAM buys speed, not feasibility. At the full size, private, sovereign web search becomes practical for a team, a company or a country.
| Index | Pages | On disk | RAM | Runs on |
|---|---|---|---|---|
| Small | 100 million | 300 GB | 32 GB | a Hetzner server, about 30 € a month |
| Medium | 300 million | 1 TB | 128 GB | a Hetzner auction server, about 130 € a month |
| Full | 1 billion | 3 TB | 384 GB | one server; the whole index downloads in under a day |
Medium and full are the next milestones. Small is here now.
Available today
Here it is. Help yourself.
The first version of the index, with 100 million pages, together with the source code to deploy it and to extend it with your own content. A hosted demo is live at websear.ch and connects to any agent via MCP.
Connect an agent
Add the hosted endpoint to your MCP client. The agent gets one tool, search, which returns the URL of the authoritative source with the minimum context needed to choose.
Your own deployment exposes the same endpoint on your own host.
{
"mcpServers": {
"cytir": { "url": "https://mcp.websear.ch/mcp" }
}
}Who's behind it
Two stewards, one commons.
Cytir is a collaboration between impactIA, a Geneva-based foundation, and Vector8, a European AI consultancy. Code and index ship under Apache-2.0.