Connect NotchBio MCP
NotchBio MCP brings accessible RNA-seq projects, recorded analyses, and supported outputs into an AI assistant. This guide covers the initial read-only preview. Availability is shown in lab Settings → MCP & AI apps.
Before connecting
Section titled “Before connecting”Sign into your NotchBio lab account first. Open Settings → MCP & AI apps and check whether setup is enabled. The published server URL appears there when available. Copy it exactly; do not use the lab website address or an example endpoint.
Your AI client must support a remote Streamable HTTP MCP server and the configured OAuth flow. Custom connections may require developer mode, a compatible account, workspace permission, or administrator approval. Client-side MCP support is not proof of a verified NotchBio sign-in. Follow your client’s current instructions if its interface differs.
ChatGPT
Section titled “ChatGPT”On ChatGPT’s web interface, open Plugins, use the plus button, and choose Add custom MCP server. Enter a name and paste the URL from NotchBio. Configure OAuth sign-in, review the permissions, and create the plugin. Install it, open a new chat, and select NotchBio with @.
Some accounts or managed workspaces restrict custom servers or expose a developer-mode entry point. Check the official ChatGPT custom-server guide rather than assuming every account has the same controls.
Claude
Section titled “Claude”In Customize → Connectors, add a custom connector using the copied URL. Complete the sign-in flow and enable the connector in your conversation. For managed workspaces, an owner may need to add the connector first.
Authentication choices and organization controls are described in Claude’s remote connector instructions.
Cursor
Section titled “Cursor”Use Cursor’s MCP customization settings to add a remote server. You can also add a named entry to .cursor/mcp.json, using the actual URL from NotchBio:
{ "mcpServers": { "notchbio": { "url": "PASTE_THE_HTTPS_MCP_URL_FROM_NOTCHBIO_SETTINGS" } }}Replace the placeholder before connecting. Complete OAuth when prompted, check the discovered tools, and enable the tools you intend to use in Agent. See the official Cursor MCP documentation for current configuration and sign-in support.
VS Code
Section titled “VS Code”Enable the AI features you use, then open the Command Palette and run MCP: Add Server. Choose HTTP, paste the server URL, and give the entry an identifiable name such as notchbio. Select the appropriate personal or workspace destination.
Review trust and authorization prompts, complete browser sign-in, and verify the tools in your agent chat. See VS Code’s MCP setup guide for harness and configuration requirements.
Gemini CLI
Section titled “Gemini CLI”This path is for Gemini CLI, not the consumer Gemini web or mobile chat app. Add an entry to the CLI’s settings.json:
{ "mcpServers": { "notchbio": { "httpUrl": "PASTE_THE_HTTPS_MCP_URL_FROM_NOTCHBIO_SETTINGS" } }}Replace the placeholder, restart the CLI, and inspect the connection with /mcp. Use /mcp auth notchbio to complete browser authorization. OAuth discovery and the local callback must be supported by the NotchBio authorization provider. A headless session may not support this browser flow.
See the Gemini CLI MCP documentation for its current transport and authentication requirements.
Verify the first conversation
Section titled “Verify the first conversation”Start with:
Read the NotchBio agent guide, list my projects, and ask which project and analysis I want to review. Check your available tools before suggesting actions.
Confirm that the assistant identifies the right project and analysis. Ask it to retain the exact comparison, including the numerator and baseline, before interpreting a result. It should read the complete relevant handbook pages and state when a tool or output is unavailable.
Useful requests include explaining recorded QC/trimming, listing analysis outputs, finding a result table, and drafting methods notes with explicit missing details. Verify scientific claims and methods against original records before publication.
Preview capabilities
Section titled “Preview capabilities”The current tools support project/run inspection, recorded QC and preprocessing, quantification summaries, comparison metadata, archive metadata assessment, bounded text reads, complete-file download links, supported raster previews and figure rendering, and agent documentation.
The preview cannot start, change, cancel, or retry an analysis; import archive files; or upload local files. Use the lab app for those actions. The assistant can explain saved defaults and help plan questions, but cannot claim that it applied a new experiment design.
Supported images may appear inline, and files can be saved using expiring download links where the host supports them. The remote service cannot silently save files to a folder on your device. A preview or snippet is not the complete output.
Permissions and connection management
Section titled “Permissions and connection management”Custom charts and further tests
Section titled “Custom charts and further tests”You can ask the assistant to discuss the returned results and, when the host has data-analysis tools, replot a complete downloaded table. NotchBio’s MCP renderer is limited to volcano, MA, PCA, and heatmap figures. Additional visualizations or statistical code run in the host’s environment, not through an arbitrary-code execution tool on NotchBio. Starting or rerunning NotchBio analyses from chat is not implemented in the current preview.
MCP uses your NotchBio identity and current project access. The scopes granted to the AI connection can further limit file downloads or log access. A connector does not grant access to somebody else’s project or facility administration.
Sign in only through the authorization page. Do not place passwords, API keys, access tokens, or private download links in a public conversation or shared configuration. Disconnect in the AI app and revoke the authorization with the identity provider when available. The lab setup section provides connection instructions; it does not claim to list or revoke every client session.
Troubleshooting
Section titled “Troubleshooting”| Symptom | What to check |
|---|---|
| Setup is not enabled | The remote-service rollout has not been published. Wait for a server URL to appear in lab settings. |
| No custom-server option | Check your AI client’s account, workspace, developer-mode and administrator permissions. |
| URL or discovery fails | Copy the exact published HTTPS URL, including /mcp. A running lab website is a separate service. |
| Browser sign-in fails | Confirm the client’s OAuth discovery and redirect flow are enabled by the provider. Do not bypass sign-in with a browser session token. |
| A project is missing | Check that the same NotchBio account has current access and that the connection is not restricted to other projects. |
| File access needs permission | Reauthorize through the client’s sign-in flow with the required file/export access. |
| A download link expired | Ask for a new link and save it through the host’s download interface. |
| The assistant offers to start a run | Ask it to check its actual tools. The current preview cannot execute analysis actions. |
For a concise walkthrough, read the MCP setup tutorial. For the product overview and illustrative conversations, visit NotchBio MCP.
Tool reference
Section titled “Tool reference”See the complete MCP tool reference for all available functions, input parameters, permissions and limitations. Connected assistants can discover and read the same reference through nb_list_documentation and nb_read_documentation.