# UDI, the Universal Data Index

> UDI, from CellType, searches what public biomedical datasets measured, not just what they are called. It indexes
> 1.9M+ public studies from 67 sources and 640K+ signatures of what changed inside their data, plus 41M+ papers,
> clinical trials and drug approvals. Each answer links to its source. You can use it in a browser or from your AI
> assistant (Claude, ChatGPT, Codex, Cursor or any MCP client). Free during the beta.

Last updated: 2026-11-03

## What UDI does

Most dataset search matches titles and descriptions. UDI also searches the results inside public studies: which genes
went up or down, in which comparison, and by how much. So a question like "Where does SPP1 go up in IPF lung?" returns
the datasets where it was measured to go up, with the fold change and a link to each GEO record. Without UDI, you or
your AI would have to download and reanalyze each dataset, which takes days. With UDI it is one search.

## What is in it

- 1.9M+ public studies from 67 sources, including GEO, SRA, ArrayExpress, PRIDE and CELLxGENE
- 640K+ signatures of what the data measured: what changed, and by how much
- 230K+ single-cell studies, with tissues, cell types and sample counts
- 41M+ papers, linked to their datasets
- 600K+ clinical trials with results, plus FDA and EMA approvals, labels and safety records
- Genetics: fine-mapped GWAS and 18M+ QTL links

## Who uses it, and what they ask

- Academic and translational scientists: "Find human single-cell RNA-seq of IPF lung with at least 10 donors."
- Target discovery: "What human genetic evidence links IL23R to psoriasis?"
- Biomarker teams: "Does this 12-gene psoriasis signature respond to treatment?"
- Clinical development: "Why did the phase 2 BEACON-IPF trial of bexotegrast stop?"
- Competitive intelligence and BD: "Who is developing IL-23 drugs for psoriasis, and how far along are they?"
- Safety and regulatory: "What is in the tofacitinib boxed warning?"

## How well it finds the right dataset

On 443 real dataset requests drawn from 2026 papers, UDI put a right dataset in its top 10 for 92% of requests,
against 73% for an AI with web search. EBI Search and GEO search were tested too.

## How to use it

- In a browser: https://www.celltype.com/udi (try a search on the page)
- From your AI: connect the MCP server https://udi.celltype.com/mcp and sign in once with a free CellType Cloud
  account. Read-only, no key to copy. Setup steps for each client: https://www.celltype.com/udi/llms.txt
- Price: free during the beta, 10,000 calls a month, no card.
- Your own data: Private UDI puts your private data beside public data in one search, inside your environment:
  https://www.celltype.com/udi/private

## Links

- [UDI page](https://www.celltype.com/udi): search, examples and benchmark
- [Setup for AI assistants](https://www.celltype.com/udi/llms.txt): Claude, ChatGPT, Codex, Cursor and other MCP clients
- [Private UDI](https://www.celltype.com/udi/private): your data and public data in one search
- [CellType](https://www.celltype.com): the company behind UDI
- Contact: udi@celltype.com

---

# UDI setup (for people and AI agents)

UDI, the Universal Data Index from CellType, is a free search index of public biomedical data: datasets
(GEO, SRA, ArrayExpress, PRIDE, CELLxGENE and more) and measured results inside them, PubMed papers, clinical
trials, FDA and EU approvals, labels, safety reports, patents and genetic evidence, each with its source.
Your AI agent uses it through MCP.

- MCP server (Streamable HTTP): https://udi.celltype.com/mcp
- Official MCP registry: [CellType UDI](https://registry.modelcontextprotocol.io/v0.1/servers/io.github.celltype%2Fudi/versions/latest)
  (`io.github.celltype/udi`). This is a hosted MCP server; no Python package is needed.
- Sign-in: OAuth with your own free CellType Cloud account (Google, GitHub, or an email code); read-only scope `udi.read`.
  There is no shared key to copy. Until you sign in, the client shows UDI as "needs authentication" with no tools.
- In a browser, without an agent: https://cloud.celltype.com/udi
- Help: udi@celltype.com

Beta testers: if your invite gives a different address, use that address instead.

## If you are an AI agent setting this up for your user

1. Find which client you are running in below and run or show its connect step.
2. Sign-in needs your user: they open the sign-in link in a browser once. You cannot do it for them.
3. Most clients load MCP servers only when a chat starts: after connecting, the user starts a new chat.
4. Check it works with: "Use UDI to tell me what it has on PCSK9." (UDI's `udi_coverage` tool answers.)

## Claude Code

```bash
claude mcp add --transport http --scope user celltype-udi https://udi.celltype.com/mcp
```

`--scope user` makes UDI available in every project. Then start a new chat (or restart Claude Code), type `/mcp`,
pick `celltype-udi`, choose **Authenticate**, and sign in in the browser window. `claude mcp login` does not work
from an agent's non-interactive shell; use `/mcp` inside a chat.

## Claude desktop app and claude.ai

1. Settings → Connectors → **Add custom connector**.
2. Name `UDI`, URL `https://udi.celltype.com/mcp`, **Add**.
3. Click **Connect** next to UDI and sign in. In a chat, open the tools menu (sliders icon) and switch UDI on.

On Team and Enterprise plans an owner may need to add the connector first.

## ChatGPT

Needs a plan that allows custom apps (Plus, Pro, Business, Enterprise, Edu).

1. Settings → Apps & Connectors → Advanced settings: turn on **Developer mode**.
2. Apps & Connectors → **Create**. Name `UDI`, MCP server URL `https://udi.celltype.com/mcp`, Authentication **OAuth**.
3. Confirm and **Create**; sign in on the UDI page that opens.
4. In each new chat: **+** → Developer mode (or More → UDI) → select UDI.

## Codex (CLI and IDE extension)

```bash
codex mcp add udi --url https://udi.celltype.com/mcp
codex mcp login udi
```

Or in `~/.codex/config.toml`:

```toml
[mcp_servers.udi]
url = "https://udi.celltype.com/mcp"
```

Check with `codex mcp list`, then `/mcp` inside Codex.

## Cursor

`~/.cursor/mcp.json` (all projects) or `.cursor/mcp.json` (one project):

```json
{
  "mcpServers": {
    "udi": { "url": "https://udi.celltype.com/mcp" }
  }
}
```

Then Cursor Settings → MCP → click **Needs login** next to `udi` and sign in.

## Any other MCP client

Connect to `https://udi.celltype.com/mcp` (Streamable HTTP) and sign in with OAuth. The server publishes its sign-in
details at https://udi.celltype.com/.well-known/oauth-protected-resource/mcp (dynamic client registration and PKCE).

## Headless agents and scripts (no browser)

A headless agent uses a read-only service token issued by CellType (ask udi@celltype.com) as
`Authorization: Bearer <token>`, on MCP or the REST API:

```bash
curl https://udi.celltype.com/udi/tools/call -H "Authorization: Bearer $UDI_TOKEN" \
  -H 'Content-Type: application/json' -d '{"name": "udi_coverage", "arguments": {"entity": "PCSK9"}}'
```

- Search: `POST https://udi.celltype.com/udi/search` with `{"query": "...", "species": "human", "limit": 10}`
- List tools: `GET https://udi.celltype.com/udi/tools`
- Python and other clients: connect any MCP client to `https://udi.celltype.com/mcp`, or call the HTTP API above with `UDI_TOKEN`.
- CellType team with a CellData checkout: `udi login` (prints a sign-in URL; works from any shell), then
  `udi --hosted <command>`; `udi doctor` shows what is set up.

## First questions to try

1. Use UDI: who else is developing TL1A drugs, and how far along are they?
2. Use UDI: is there genetic evidence that PCSK9 matters in coronary artery disease, and what did evolocumab show?
3. Use UDI to find human single-cell RNA-seq of idiopathic pulmonary fibrosis lung, check the top three studies
   for at least 10 donors, and tell me which one to download first.

Say "use UDI" the first time in a chat. Every answer lists ready next calls; ask your agent to follow them.

## Feedback

Something wrong, missing or confusing? Tell your agent "tell UDI that …" with what you asked and what was wrong. If
your agent has no UDI feedback tool, it should email udi@celltype.com with your question and UDI's answer.

More: https://www.celltype.com/udi