Click on "Deploy Server".
Wait a few minutes for the server to deploy. Once ready, it will show a "Started" state.
In the chat, type @ followed by the MCP server name and your instructions, e.g., "@propaths-mcp Search ProPaths for ATXN3 and summarize its strongest mechanistic interaction."
That's it! The server will respond to your query, and you can continue using it as needed.
Here is a step-by-step guide with screenshots.
An MCP server that exposes the ProPaths
verified protein-interactome as read-only tools for AI agents. It is a thin
client over the public ProPaths API (https://www.propaths.net), so every tool
returns exactly the API's JSON. No account, no API key.
ProPaths reads a protein's primary literature and returns a verified graph of typed, directed, mechanistic interactions plus a pathway ontology. One protein (ATXN3) is fully mapped today.
Add this to your MCP client config. uvx fetches and runs the server; nothing
to clone or install.
{
"mcpServers": {
"propaths": {
"command": "uvx",
"args": ["propaths-mcp"]
}
}
}Then ask, e.g., "search ProPaths for SCA3 and summarize its strongest
mechanistic interaction." The agent will call search_proteins then
get_protein, and drill in with get_interaction.
Prefer the raw API? It is public and keyless:
curl https://www.propaths.net/api/protein/ATXN3
curl 'https://www.propaths.net/api/search?q=SCA3'Related MCP server: dotdog
Tool | What it does |
| Find a protein by symbol, alias, or name (start here) |
| Compact interactome overview (the main entry point) |
| One interaction's full mechanism + evidence |
| The interaction(s) between two proteins, in one call |
| Filtered/sorted headline rows |
| The controlled vocabulary (edge kinds, types, directions) + counts |
| A pathway node with its lineage and member interactions |
| The full pathway scaffold (resolves pathway ids to names) |
| The best-evidenced interactions |
| Export a protein's network as TSV / SIF / GraphML (Cytoscape, networkx) |
| The graph vocabulary + how to use the tools (offline) |
Also exposed as MCP resources (propaths://schema, propaths://interaction-types,
propaths://pathways/tree, and the propaths://protein/{symbol} template) and
prompts (profile-protein, strongest-evidence, explain-pathway).
All tools are read-only and idempotent.
Env var | Default | Purpose |
|
| API base URL. Point at |
pip install propaths-mcp
propaths-mcp # runs the stdio server
# or: python -m propaths_mcpBefore it is published, you can run straight from the repo:
uvx --from git+https://github.com/Tahsin-Kazi/propaths-mcp propaths-mcpRead-only and public; reads are rate-limited per client. Write/enrichment access and a hosted MCP are gated. Get in touch.
Errors are graceful: a missing protein returns {"error": "...", "status": 404};
an unreachable API raises with a hint.
Docs: https://www.propaths.net/quick-start · License: MIT
Explain the graph's vocabulary and how to use these tools: edge kinds, interaction types, direction semantics, orientation, and the recommended call flow. Static guidance that works even if the API is unreachable.
Inputs: no named parameters.
Export a protein's interactome as text for external tools.
format = tsv (edge list / spreadsheet), sif or graphml (Cytoscape, networkx, igraph, Gephi). Edges carry their biological orientation (an upstream partner points into the protein), so the graph is directed.
Inputs: symbol (required), format. Returns: result
A curated entry point: the best-evidenced interactions, ranked by supporting evidence. Good for a quick, strong overview.
Inputs: no named parameters.
One interaction's full enriched record by id.
Includes mechanism prose, direction, per-function effects, kinetics, and
evidence. Pass query (a participating symbol) to orient source/target so
the query protein reads as the source.
Inputs: interaction_id (required), query.
The interaction(s) between two named proteins, oriented from a.
Use this for "what does A do to B" in one call, instead of pulling get_protein and scanning for the partner.
Inputs: a (required), b (required).
A single pathway by id: the node, its ancestors and children, and the interactions placed in it. The pathway-first way into the graph.
Inputs: pathway_id (required).
The full pathway scaffold as a flat node list.
Assemble the tree client-side via each node's parentId. Use this to resolve the canonicalPathwayId values returned on interactions into human-readable pathway names.
Inputs: no named parameters.
Overview map for one protein, the main entry point.
Returns a COMPACT overview sized for an agent: protein metadata, counts, the top-level pathway roots, and one headline row per interaction (id, oriented source/target, kind, direction, type, a one-line supportSummary, resolved pathway names, and function/evidence counts). It deliberately omits the heavy per-edge mechanism and evidence prose (the full protein page is ~700K tokens for a hub protein).
Drill into any row by id for full depth: get_interaction(id) for an edge's mechanism + kinetics + evidence.
Inputs: symbol (required).
Filtered, sorted, headline-only list of a protein's interactions.
kind = activates|inhibits|binds|regulates; type = direct|indirect; pathway = a pathway-name substring; min_evidence = minimum supporting papers; sort = evidence|functions|partner. Lighter than get_protein; use it for targeted questions ("best-evidenced inhibitory edges in ERAD"), then drill in with get_interaction(id).
Inputs: symbol (required), kind, sort, type, +4 more.
The controlled vocabulary with plain-language meanings and live counts: edge kinds, interaction types, directions, and the mechanisms present in the graph. Call this before filtering so you use valid values.
Inputs: no named parameters.
Search proteins by symbol prefix (case-insensitive).
Returns matching symbols with HGNC id and description. Use this first to resolve a protein of interest, then call get_protein.
Inputs: q (required), limit.
Recent tool additions, removals, and schema changes observed during successful MCP inspections.
Loading tool schema changelog…
Scored across 11 tools
Each tool has a clearly distinct purpose: search_proteins for finding proteins, get_protein for overview, get_interaction for deep detail, list_interactions for filtered lists, get_pathway/get_pathway_tree for pathway navigation, export_network for output, and describe_schema for guidance. Even closely related tools like get_interaction_between and get_interaction are differentiated by their argument (pair vs. ID) and described usage.
All tool names follow a consistent get_/list_/search_/export_/describe_ + noun pattern, all in snake_case. The verb clearly indicates the action (get for single resource, list for collection, search for prefix query, export for output, describe for metadata) and the noun identifies the resource (protein, interaction, pathway, network, schema).
11 tools is well within the recommended 3-15 range and each tool addresses a specific need for exploring a biological interaction graph. The count feels neither thin nor bloated—there are distinct entry points, filtering utilities, vocabulary support, and export functionality without redundant overlaps.
The tool surface covers the full read-only lifecycle of the domain: find a protein (search_proteins), get overview (get_protein), drill into details (get_interaction), filter (list_interactions), navigate pathways (get_pathway, get_pathway_tree), discover vocabularies (list_interaction_types), get curated highlights (get_highlights), and export (export_network). The inclusion of describe_schema ensures agents can self-orient, and there are no obvious dead ends—every tool's output either stands alone or provides IDs that are directly consumed by other tools.
Read-only MCP tools for AI agent discovery, structured resources, and NIULAI information.
Reactome MCP — open biological pathway knowledge-base.
Read-only search of 2,189 AI agent tools (MCP servers, x402 APIs) with daily automated trust scores.
Public read-only MCP access to digital learning concepts, semantic relationships, and citations.
