Use COMPASS with your agent

Search all 5,138,521 clusters the pipeline produced, or only the 255,213 RNA clusters the website highlights, explore their biology, and retrieve sequences and alignments from the tools you already use.

MCP server URL

https://compass.arcinstitute.org/mcp

Read-only accessGoogle sign-inRelease atlas_release_2026-10-r4Two data scopes

Connect your client

Add the server, then sign in with Google when prompted. Browsing the atlas needs no account.

Claude Code

Run this command in your terminal:

claude mcp add --transport http ncrna-atlas https://compass.arcinstitute.org/mcp

In Claude Code, enter /mcp, choose ncrna-atlas, and follow the sign-in flow. Claude Code guide.

Claude Desktop or claude.ai

Add a custom connector, paste the server URL above, and sign in with Google when prompted. If asked for an OAuth client registration method, choose automatic registration.

Claude connector guide.

ChatGPT

Open Plugins → + → Add custom MCP server. Enter the server URL, choose OAuth, and sign in with Google. Availability depends on your account and workspace permissions.

ChatGPT connection guide.

Cursor

Add this configuration to .cursor/mcp.json in your project or ~/.cursor/mcp.json for all projects, then sign in from the MCP settings:

{
  "mcpServers": {
    "ncrna-atlas": {
      "url": "https://compass.arcinstitute.org/mcp"
    }
  }
}

Cursor MCP guide.

Scripts and other MCP clients

Use the same URL with an MCP client that supports Streamable HTTP and OAuth. For a script or an unattended workflow, contact the lab about token access.

Authenticated requests use Authorization: Bearer <token>. Keep tokens private.

A few things to ask

Start with a biological question. Your agent can look up exact annotation terms and link the results back to the atlas.

Which clusters with strong covariation sit next to a CRISPR-associated gene in Bacteroidota?
List the Rfam families whose top hit covers more than half of a cluster’s members, with cluster counts.
Among the highlighted clusters the website shows, how many have a pseudoknot and a terminator-like 3′ end? Break it down by phylum.
Open cluster 217088: show its structure, neighboring genes and taxonomy, and get me its alignment.
How many clusters outside the master table have a T-box hit? Show the species of one.

Tool reference

The server uses the Model Context Protocol. Agents should start with describe_data to learn the schema and query language. Every tool takes a scope argument, all (default) or highlighted; the two scopes are described below.

Browse the seven tools

describe_data: Tables, key columns, the condition language, and request limits.
describe_columns: A dictionary of every cluster column, by group or name.
suggest_terms: Exact Pfam, Foldseek fold, taxon, and Rfam names for searches.
search_clusters: Free text with gene, taxon, Rfam, numeric, and motif conditions.
search_ids: Matching cluster IDs, paginated for larger result sets.
query_sql: Read-only SQL SELECT queries over clusters, partners, taxa, Rfam, and structure tables.
get_cluster: Cluster statistics, consensus structure, covariation, neighboring genes, taxonomy, Rfam hits, member genomic context, and download links.

Two data scopes

Every tool takes a scope argument; without one it searches everything. describe_data states each scope's gaps, tables and limits in detail; get_cluster notes when a cluster is outside the master table or has no page.

all (default): Data release atlas_full_2026-10-f1: every cluster the pipeline produced (5,138,521 clusters), the highlighted ones among them. Most never entered the master table, so their motif, consensus structure, covariation and enrichment columns are null; in_master, has_structure and highlighted say what a row has, and members, loci, neighboring genes, taxonomy, Rfam hits and terminator scans exist for all of them. The per-member tables (members, loci, genes, terminators) are large Parquet tables partitioned by bucket = cluster_id // 10000: filter on the bucket and the cluster id, or a query runs into the time limit. Coordinates are on the source contig. Web pages exist only for highlighted clusters.
highlighted: Data release atlas_release_2026-10-r4: the 255,213 highlighted clusters the website shows, each with a web page and nearly all with a drawing and alignment viewer; per-member genomic context in the website's frame (positions relative to the RNA).

Access and limits

Up to 120 tool calls per minute per account, per server instance; in the all scope also 2,000 tool calls per account per day. A busy server asks your agent to retry. The edge also limits any one IP address to 600 requests per minute on /mcp and answers 429 above that.
Searches return up to 200 rows per page; ID lists up to 10,000 per page. SQL is read-only and returns up to 1,000 rows; it stops after 30 seconds in the highlighted scope and 120 seconds in the all scope, where 8 queries run at once. Results are capped at 100,000 characters and report truncated: true when shortened.
Member sequences (FASTA) and CaCoFold motif alignments (Stockholm) are available per cluster through the links returned by get_cluster (/dl/all/<id>/… in the all scope, /dl/<id>/… in the highlighted scope), using the same token.

Data, citation and privacy

COMPASS, from the Hsu lab at Arc Institute, contains candidate structured RNAs mined from bacterial and archaeal intergenic regions with Evo 2. Cite the preprint (Sun, Hudelmaier et al., 2026) and the release you used, atlas_release_2026-10-r4 for the highlighted clusters or the whole release named above; the About page has the full citation and BibTeX.

Tool calls and sign-ins are logged with your account’s email. Read the privacy notice for details.

Questions or access requests: the Hsu lab at Arc Institute (hsulab@arcinstitute.org).