GeneFoundry 生物医学数据库 MCP 路由器

GeneFoundry 提供生物医学数据库的 MCP 路由器与资源目录,让用户通过 Claude、ChatGPT 等 AI 聊天工具直接检索 gnomAD、ClinVar、HPO 等数据源,用于科研查询与工作流验证。

  • by berntpopp
  • TypeScript
  • 本地部署
  • 1 星标
  • 更新于
在 GitHub 查看 berntpopp/genefoundry

使用说明

GeneFoundry

Biomedical databases in your AI chat.

The website behind genefoundry.org, the source directory, setup guides and research workflows for the GeneFoundry MCP router. The backend that federates the *-link fleet lives in genefoundry-router.

Development

Use Node 24 and the committed lockfile.

npm ci
npm run dev -- --host 127.0.0.1

Development uses the same server renderer and route resolver as the production build, with styles loaded before hydration. Generated output is excluded from the development watcher. Vite defaults to /. Use VITE_BASE_URL=/genefoundry/ npm run dev for the mirror layout. This website presents the source directory and connection guides; it does not run the MCP gateway or authenticate a client.

Production and verification

npm run format:check
npm run type-check
npm run lint:check
npm test
npx playwright install chromium
npm run build:matrix
npm run test:e2e:all

The matrix builds / and /genefoundry/ sequentially, retaining artifacts in .build/artifacts/. The complete browser suite exercises Vite, both static artifacts and an isolated Docker/nginx container. Docker must be available for the HTTP checks. Reports and caches are ignored by Git.

To build one deployment explicitly:

VITE_BASE_URL=/ npm run build
# Or: VITE_BASE_URL=/genefoundry/ npm run build

Each build prerenders all registered pages, validates content and references, generates the service worker from final HTML, then creates compression sidecars. The runtime image serves only the public artifact. The edge proxy separately owns MCP and OAuth routes; see deployment notes.

Using GeneFoundry

The site provides setup guides for seven clients, with Claude and ChatGPT first, followed by Claude Code, Codex CLI, Cursor, Gemini CLI and VS Code. Add the hosted endpoint https://genefoundry.org/mcp using your client’s instructions and complete browser sign-in when requested.

Two research workflows include copyable prompts and real captured results:

  • HNF1B: gnomAD gene constraint alongside five source-linked ClinVar variant records.
  • Renal cysts and diabetes: complete HPO gene-list comparison, 13 shared genes, and the returned HNF1B ClinGen curation.

The pages include retrieval dates, source versions where returned, original record links and the exact executed calls. They are for research, not clinical decision support.

Content and design

The catalog is maintained inventory, not a live health measurement. Client guides use three distinct states: documentation-only provides browser instructions without a code snippet; documented provides reviewed code/configuration with no execution claim; verified additionally requires sanitized execution evidence. The current release has two documentation-only browser guides, five documented coding-client recipes and no verified setup recipes. The two verified workflow results were obtained through an existing Claude Code connection; they do not establish that a fresh client add/login recipe was tested.

Archivo headings and Source Sans 3 body text are self-hosted with their font licenses. All page kinds share the same 1200 px outer alignment and use responsive content columns rather than a narrow page-wide reading strip.

See DESIGN.md, the usage revision and the verification record for the final implementation. The original design specification, execution plan and shared contracts retain planning context; the later user-directed revision supersedes their initial illustrative workflows and documentation-only recipe restriction.

Generated audits, screenshots and local design-tool state are ignored. Source files, font licenses, sanitized execution evidence and the build-only verification ledger remain versioned. The runtime artifact excludes private evidence paths and account/session data.

Lighthouse across all pages

After building the matrix, serve the production Docker image at http://127.0.0.1:4182/ and run:

npm run audit:lighthouse

This runs Lighthouse 13.4.0 sequentially on all indexable routes with mobile throttling. Reports and a route summary stay under ignored .build/lighthouse/. Pass -- --url http://127.0.0.1:PORT/ to audit another production preview. Development-server performance is not representative of the compressed production build. The /genefoundry/ mirror and error pages intentionally use noindex; audit the root deployment for production SEO.