Mosaic Pre-Clinical Intelligence

io.github.sourabhnk由 GitHub 帳號 sourabhnk 發布v2.2.1更新於 Oct 25, 2026

Pre-clinical drug-target intelligence: 60 curated targets, 1,969 counted partner genes, 11 tools

發布方自己的服務：[PREPROD] Clinical.ly Research Suite

已驗證STDIO僅桌面資料庫 AI 與機器學習

服務詳細資訊

識別碼: io.github.sourabhnk/mosaic-mcp
傳輸: STDIO
執行環境: Desktop only
信任: Verified
發布於: Sep 16, 2026

安全提示

MCP 服務會接收對話中的工具呼叫請求，並可能執行外部操作。啟用前請先了解該服務能做什麼，並只授予它所需的權限。

註冊中心已驗證此條目所屬的 GitHub 帳號 sourabhnk。這並不表明該帳號就是產品方。

SourceWeft 的檢查只涵蓋條目是否完整、端點是否回應；不審查伺服器的程式碼，也不審查其工具的行為。

在 README 中新增徽章

向讀者展示該 MCP 服務已收錄於 SourceWeft，徽章連結到本頁。

[![Listed on SourceWeft](https://sourceweft.com/badge/mcp/io.github.sourabhnk%2Fmosaic-mcp.svg)](https://sourceweft.com/mcp/io.github.sourabhnk%2Fmosaic-mcp?utm_source=readme&utm_medium=badge)

概覽

AI 產生的概覽

本機 MCP 伺服器，以你自己的 PostgreSQL Mosaic 資料庫回答臨床前藥物標靶問題，內含 5 個免費與 6 個 Pro 工具。

可以做什麼: 查詢標靶檔案、具實測活性的化合物、相關論文與 AlphaFold 結構，每項都附覆蓋狀態，缺漏資料不會顯示為零。對任意人類基因符號檢視覆蓋網格，身分與結構即時解析。啟用 Pro 後可查看臨床試驗、實驗驗證先例、通路背景及標靶的一跳網路。評估合成致死空白區與候選抗藥旁路標靶。免費呼叫最多回傳 10 筆化合物或論文，Pro 最多 50 筆。
適用情境: 當你已擁有含 Mosaic schema 的 PostgreSQL 執行個體，並希望助理透過 stdio 在本機查詢藥物標靶情報時使用。它並非 Mosaic 精選資料的來源；README 指出這類資料需改用另一個託管伺服器。
執行需求: 本機 Python 執行環境（pip、uvx 或 python -m），以及 DATABASE_URL 中的 PostgreSQL 或 Neon 連線字串，須以 /mosaic_db 結尾並包含 Mosaic schema。可選 MOSAIC_TIER 與 MOSAIC_API_KEY。僅支援 stdio，本套件未實作遠端傳輸。

安裝前請注意

DATABASE_URL 是含憑證的機密連線字串，且為必填。MOSAIC_API_KEY 不在本機驗證：stdio 模式下任何值都視為 Pro，因此層級限制不會生效。本套件只提供查詢，不含資料，也不提供共用唯讀執行個體。檔案工具僅在託管伺服器上可用。

安裝

在 SourceWeft 中

開啟儀表板中的 Mosaic Pre-Clinical Intelligence，將其新增到工作區。
為需要使用其工具的對話啟用該服務。

Desktop only，透過 STDIO。 STDIO 服務會啟動本機處理程序，因此需要 SourceWeft 桌面主機。

其他 MCP 客戶端

參照儲存庫中的啟動說明。

README

mosaic-mcp

Pre-clinical drug-target intelligence as an MCP server: target profiles, compounds, literature, structure, clinical pipeline, validation precedent and whitespace analysis — 11 tools, 5 free and 6 Pro. Every count carries a coverage state, so "we did not look" never renders as zero.

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, with the Mosaic schema, and it serves the tools over it. There is no public read-only credential for Mosaic's knowledge graph.

If you want Mosaic's curated knowledge graph — 60 oncology targets, ~128K compounds linked to them, 57,896 papers and 9,062 clinical trials (counted 2026-10-07), refreshed monthly — use the hosted server. It needs no database and no local install:

Claude.ai connector (OAuth, Streamable HTTP): https://mcp.getmosaic.dev/mcp
Remote MCP with an API key (SSE): https://mcp.getmosaic.dev/sse
Sign in, API keys and on-demand dossiers for any human gene: https://getmosaic.dev

Quick start (self-hosted)

With Claude Desktop

Add to your claude_desktop_config.json:

json

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

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

With Claude Code

Standalone

bash

export DATABASE_URL="postgresql://..."   # your own Postgresmosaic-mcp                               # stdio transport

export DATABASE_URL="postgresql://..."   # your own Postgresmosaic-mcp                               # stdio transport

uvx mosaic-mcp and python -m mosaic_mcp.server work too. This package is stdio-only: --transport sse exits with NotImplementedError, because remote transport is the hosted server above.

Tools

Free (5)

Tool	What it returns
`mosaic_get_target_profile`	The target dossier: biology, approved and clinical drugs, disease associations ranked by Open Targets, validation evidence, organisations, scores — each axis with its coverage state
`mosaic_get_target_compounds`	Compounds with measured activity on the target; `sort` by potency (default) or by max clinical phase
`mosaic_get_target_papers`	Papers about the target, matched by NCBI Gene ID (PubTator3) where your database has the annotations, up to the declared corpus cutoff
`mosaic_get_target_structure`	AlphaFold structural snapshot and pockets, with a caution when a pocket is lined by low-confidence residues
`mosaic_get_coverage_grid`	A coverage grid for any human gene symbol: identity (HGNC) and structure (AlphaFold DB) resolve live; the other axes are marked `queued`

Free calls return up to 10 compounds or papers; Pro up to 50.

Pro (6)

Tool	What it returns
`mosaic_clinical_pipeline`	Clinical trials for compounds acting on the target (a ChEMBL mechanism link or pChEMBL ≥ 6)
`mosaic_target_validation`	Assay precedent: what has been tried, in what model system
`mosaic_pathway_context`	The pathways the target sits in
`mosaic_target_network`	The target's 1-hop neighbourhood: compounds, diseases, pathways, organisations, interacting proteins
`mosaic_synthetic_lethal_whitespace`	Partners functionally coupled to the target with little drug activity — assessed candidates and unassessed partners kept apart
`mosaic_resistance_bypass_map`	Candidate resistance-bypass and escape targets

mosaic_request_dossier and mosaic_get_dossier (on-demand dossiers) run only on the hosted server; they need its fetch and synthesis pipeline.

Configuration

Variable	Required	Description
`DATABASE_URL`	Yes	PostgreSQL connection string for your own instance; the path must name the database that holds the Mosaic schema.
`MOSAIC_TIER`	No	`pro` (or `enterprise`) enables the Pro tools in a self-hosted install.
`MOSAIC_API_KEY`	No	In stdio mode any value is treated as Pro; it is not checked locally. Keys are validated by the hosted server.

Changelog

Full history in CHANGELOG.md. Latest: 2.1.1 — the mosaic-mcp command starts again (it had failed with ImportError since 1.7.0). 2.1.0 capped mcp below 2, without which a fresh install died on import, and carried the fixes from a live audit of the hosted server.