mosaic-mcp

Pre-clinical drug discovery intelligence as an MCP server. Query 760+ drug targets, 70K+ compounds, 48K+ papers, 18K+ clinical trials, and 16K+ patents through 44 specialized tools — 16 free for discovery, 28 Pro for competitive landscapes, whitespace, and thesis-grade analysis.

Which one do you want?

This package is bring-your-own-database. It is the MCP tool layer only — it ships the queries, not the data. You point it at a PostgreSQL instance you control and it serves 44 tools over it. There is no public read-only credential for Mosaic's knowledge graph, and earlier versions of these docs implied otherwise.

If you want Mosaic's actual curated KG — 760+ targets, 70K+ compounds, 48K+ papers, 18K+ trials, 16K+ patents, kept current by a monthly refresh — use the hosted server instead. It needs no database and no local install:

• Remote MCP: https://mcp.getmosaic.dev/sse
• Sign in / API keys: https://getmosaic.dev

Quick Start (self-hosted)

pip install mosaic-mcp

With Claude Desktop

Add to your claude_desktop_config.json:

{
  "mcpServers": {
    "mosaic": {
      "command": "mosaic-mcp",
      "env": {
        "DATABASE_URL": "postgresql://USER:PASSWORD@YOUR-HOST:5432/mosaic_db?sslmode=require"
      }
    }
  }
}

With Claude Code

claude mcp add mosaic -- mosaic-mcp

Standalone

export DATABASE_URL="postgresql://..."   # your own Postgres
export MOSAIC_API_KEY="msk_..."          # Optional — for Pro tools
mosaic-mcp                               # stdio transport

This package is stdio-only. --transport sse exits with NotImplementedError; remote transport is served by the hosted endpoint above, not by this package. The flag was previously documented as working.

Tools

Free Tier (16 tools) — discovery + workspace

Pro Tier (28 additional tools) — analysis + whitespace

The committed-verdict layer. Includes the whitespace and differentiation tools: mosaic_synthetic_lethal_whitespace, mosaic_modality_gaps, mosaic_resistance_bypass_map, mosaic_find_undruggable_targets, mosaic_talent_migration, mosaic_emerging_signals, and mosaic_assess_druggability.

Plus the full analysis set: competitive landscape, pathway context, compound selectivity, indication landscape, target validation, clinical pipeline, compound analogs, target comparison, similar-target search, opportunity finding, organization portfolio, target network, mechanism of action, evidence maps, relation search, polypharmacology, clinical trial results, FDA regulatory status, drug comparison, drug repurposing candidates, and KOL discovery.

Data Coverage

_{Counts as of 2026-07-18. The live figures are always mosaic_kg_stats; run it rather than trusting this table. Targets went 802 → 764 when duplicate and malformed rows were merged, and organizations 153,852 → 36,691 when affiliation parsing was corrected — both are the count getting more honest, not the corpus shrinking.}

Configuration

Changelog

1.2.0 — breaking change to the watchlist tools

owner_key must now be anon:<token>. It was previously free text documented as "user id, email, or anon:<token>", which meant one caller could name another and read their lists. A bare id or email is now refused.

Generate a random token, keep it, and pass the same one every time — it is what proves a list is yours. Calls that pass an email or bare id now fail with a message telling you this; they do not silently return someone else's data, which is what the old behaviour risked.

Also in 1.2.0:

• Watchlist reads are owner-scoped in the SQL, so a watchlist UUID is no longer sufficient for read or write access, and they are excluded from the response cache (a cache hit would otherwise return before the ownership check ran).
• mosaic_synthetic_lethal_whitespace now returns candidates and coupled_unassessed as two separate lists. Partners outside the curated universe carry null counts instead of zeros — an unmeasured competitor count previously read as "no competition."
• Project metadata now points at getmosaic.dev. 1.1.0 and earlier pointed at a domain that is not ours.
• Query/response layer resynced to the hosted server. This includes the protein-protein-interaction reader fix: 1.1.0 counted PPIs from the legacy target_interactions table, while the hosted server had already moved to the populated target_interactions_ext table. Driven against hosted on 2026-07-19, EGFR reports 50 interactions at confidence 0.999; the 1.1.0 reader does not read that table for the profile count.

License

Apache 2.0 for the MCP server tools. The hosted knowledge graph data requires a subscription for Pro-tier access.