Connect AI assistants (MCP)

CyteType runs a Model Context Protocol (MCP) server. Add it to an AI assistant that supports MCP, such as Claude, Claude Code or Cursor, and the assistant can open your CyteType jobs, read their annotations and the evidence behind them, query their data and compare jobs with each other. All tools are read-only: they never change your jobs.

Server URL: https://mcp.cytetype.nygen.io

The main use is synthesis across jobs. Connect several reports and ask the assistant which findings agree, which differ, and which differences are supported by marker genes, metadata, expression, differential expression or cell-cell interactions. Typical uses:

• comparing disease and control cohorts;

• checking whether an annotation reproduces across repeated runs;

• comparing similar tissues, donors, batches or time points;

• finding shared and study-specific cell populations.

Chat is the alternative inside CyteType: it uses CyteType-hosted models and saves figures and tables. The MCP connector brings the same data to the assistant you already use.

Before you start¶

• A CyteType account. You sign in with your institutional email address, as on the website (see Sign in and API keys).

• An MCP client that supports remote servers over HTTP with OAuth sign-in.

You do not need an API key: the connector uses its own sign-in.

Set up your client¶

Claude (web, desktop and mobile)¶

1. In Claude, open Customize, then Connectors (https://claude.ai/customize/connectors).

2. Click Add custom connector.

3. Enter the name CyteType and the server URL https://mcp.cytetype.nygen.io, and continue through the dialog. If it asks how to register the OAuth client, choose Register automatically: the CyteType server supports automatic (dynamic) registration.

4. Click Add, then sign in (see Sign in). If the connector shows Connect instead, click it to sign in.

5. In a conversation, click + next to the message box, choose Connectors and turn CyteType on.

On Team and Enterprise plans, an organization owner first adds the connector under Organization settings, Connectors. Each member then finds it under Customize, Connectors, clicks Connect and signs in with their own CyteType account. A connector added in Claude is available in the web app, Claude Desktop and the mobile apps.

Claude Code¶

Add the server from your terminal:

claude mcp add --transport http --scope user cytetype https://mcp.cytetype.nygen.io

--scope user makes the connector available in all your projects; without it, it is added to the current project only. Then start Claude Code, run /mcp, select cytetype and choose Authenticate. Your browser opens the CyteType sign-in page.

claude mcp list shows whether the server is connected or still needs authentication. claude mcp remove cytetype removes it, together with its stored sign-in.

Cursor¶

Add the server to ~/.cursor/mcp.json (all projects) or to .cursor/mcp.json in a project folder:

{
  "mcpServers": {
    "cytetype": {
      "url": "https://mcp.cytetype.nygen.io"
    }
  }
}

Cursor supports OAuth sign-in for remote servers: when it connects to the server, it sends you to the CyteType sign-in page. You can enable or disable the server in Cursor’s MCP settings.

Other clients¶

Any MCP client can connect if it supports:

• remote servers over Streamable HTTP;

• OAuth 2.0 sign-in with dynamic client registration and PKCE;

• a sign-in redirect to a local address on your computer (http://localhost or http://127.0.0.1, any port). Redirects to Claude and Cursor are also accepted; a client that registers any other redirect address is refused.

A client that only runs local (stdio) servers can connect through the mcp-remote bridge, which needs Node.js and handles the sign-in through your browser. In the client’s MCP configuration:

{
  "mcpServers": {
    "cytetype": {
      "command": "npx",
      "args": ["mcp-remote", "https://mcp.cytetype.nygen.io"]
    }
  }
}

Sign in¶

The first time your client connects, it opens the CyteType sign-in page (“Sign in to connect”):

1. Enter your email address and click Continue.

2. Enter the 8-digit code from the email. It is checked as soon as you have entered all eight digits, or when you click Verify.

3. The page confirms and sends you back to your client.

Signing in follows the website’s rules: you need an institutional email address, a first sign-in creates your account, and reports already shared with that address become available.

Finish signing in within 30 minutes; after that, start the connection again from your client. Your client then stays signed in while you use it: access renews automatically and expires after 6 days without use, after which you sign in again. To stop a client from using your account, remove or disconnect the connector in that client. CyteType cannot end a client’s sign-in from its side: the access a client already holds lapses on its own, within an hour, and its renewal within 6 days.

You must sign in to use the connector, also for public reports. Once signed in, you can open:

• public reports;

• your own reports;

• reports shared with you;

• reports visible to your organization, while you are a member.

These are the same rules as for the report page and Chat (see Visibility and sharing). Unlike Chat, the connector also opens jobs that are still running.

Connect jobs¶

Ask the assistant to connect a job by its report ID or report URL (https://cytetype.nygen.io/report/<job_id>). Without an ID, it can list your jobs or, in clients that support it, let you choose one. The response gives the job’s status:

It also gives the number of clusters and annotated clusters, the start of the study context and a summary per cluster.

• Up to 10 jobs can be connected at once.

• Connected jobs belong to your account and your client, so every conversation in that client sees the same list. They are forgotten after 12 hours in which no tool call uses them.

• With one job connected, the assistant does not need to name it. With several, it passes the job ID to each tool.

• Access is checked when a job is connected and again at least once a minute while it is used. If a job is deleted or your access ends, it is disconnected and the next tool call says why.

Tools¶

The tools fall into six groups. Your assistant chooses them from your request; you do not need to name them.

Session¶

Job overview and comparison¶

Expression and metadata¶

Annotation evidence¶

Quality and review¶

Targets and labels¶

Reading the results¶

• reviewerConfidence (High, Moderate or Low) comes from the independent review step. selectionScore is the annotation step’s 0 to 1 score for its final choice and is on a different scale.

• classification is catalogue (the label matches an entry in the reference cell-type hierarchy), outOfCatalog (a coherent identity without a reference entry) or unresolved (no identity could be assigned).

• A cluster without a result has annotated: false. Your own label is kept in originalLabel and is never presented as CyteType’s label.

• compare_connected_jobs matches clusters by Cell Ontology term first, then by identical L1 to L5 levels, and finally by label text alone. Rows matched by text alone are flagged textOnlyMatch; treat them as weak evidence. Clusters without an annotation or with an unresolved identity are listed as excluded.

• run_differential_expression names its method. With the job’s uploaded expression data, it runs a cell-level Wilcoxon rank-sum test on up to 5,000 cells per group, with Benjamini-Hochberg adjusted p-values; cells are treated as independent observations, so the p-values can overstate significance. Without uploaded data, it compares the percentage of cells expressing each gene, with no statistical test and no p-values.

• query_metadata accepts one SELECT from obs, with WHERE, GROUP BY, HAVING, ORDER BY, LIMIT and window functions; joins, subqueries, WITH, multiple statements and file or network functions are rejected.

Example prompts¶

• “Connect to job <job_id> and summarize its clusters with their reviewer confidence.”

• “What share of cells express CD3E, MS4A1 and LYZ in each cluster?”

• “Compare clusters 0 and 3 against clusters 1 and 5 with differential expression, and tell me which method was used.”

• “Show the evidence for cluster 2’s label, the other candidates that were considered and how the final label was chosen.”

• “How many cells per condition are in each cluster?”

• “How does IL7R expression differ between conditions within each cluster?”

• “Which clusters have Low reviewer confidence, and what did the expert reviewers flag?”

• “Connect jobs <job_id_1> and <job_id_2> and compare their cell populations. Check each job’s organism and tissue first, and treat label-only matches as tentative.”

Limits¶

Besides the limits on connected jobs and query sizes described above, two tools have hourly limits per user: cell-level differential expression with run_differential_expression (20 runs per hour) and compare_labels_cyteonto (10 requests per hour). When an hourly limit is reached, the tool says when to try again. Usage limits and quotas lists every connector limit.

Privacy¶

• Read-only. The tools cannot change your jobs, their visibility or their sharing. disconnect_job only changes the list of connected jobs.

• Your access, nothing more. The assistant can open only the jobs your account can open, and access is checked again at least once a minute while a job is in use.

• What the assistant sees. Everything a tool returns is sent to your AI client and the model provider behind it: annotations, study descriptions and metadata tags and, for jobs with uploaded data, rows of the cell metadata table (up to 200 per query) and expression statistics. How that data is stored and used is governed by your AI provider’s terms and settings, not by CyteType. Check that your data-use agreements allow it before you connect jobs with sensitive metadata.

• What CyteType sees. CyteType receives the tool calls your assistant makes, with their arguments (such as job IDs, gene lists and SQL queries), not the rest of your conversation.

See Security and privacy.