SME Library MCP
Enables AI agents to create, search, and manage subject matter expert profiles, including private workspace organization and moderated sharing to a common library.
README
SME Library MCP
A free, open-source MCP server for creating, searching, and sharing Subject Matter Expert profiles that AI agents can consult. Anyone can connect their agent, pull vetted experts from the shared library, maintain private experts in their own workspace, and contribute good ones back through a moderated promotion pipeline — or self-host the entire stack.
Licensed under the MIT License. Contributions welcome via issues and pull requests.
Runs on Vercel (Next.js + Streamable HTTP transport) with a Supabase backend.
New here? Read the User Guide — how it all works plus a complete
reference for every SME field and its value ranges (also served in-app at /guide).
See ARCHITECTURE.md for the full design and roadmap, and
docs/SME_MCP_Implementation_Checklist.pdf for the
step-by-step deployment runbook (regenerable from docs/implementation-checklist.html).
Tools
| Tool | Scope | What it does |
|---|---|---|
search_smes |
read | Full-text + tag + expertise search across library and workspace |
list_smes |
read | Browse library / workspace SMEs, ranked by quality |
get_sme |
read | Fetch one full profile |
create_sme |
write | Author a new workspace SME (versioned from v1) |
import_smes |
write | Bulk-import up to 200 SME profiles; invalid rows reported and skipped |
export_smes |
read | Export workspace SMEs as an import-ready JSON payload |
update_sme |
write | Edit a workspace SME; every edit snapshots version history |
archive_sme |
write | Soft-delete (reversible) |
clone_sme |
write | Copy a library SME into your workspace as an editable private copy |
generate_sme |
write | AI-generate a profile — dedups against the library first; quota-limited |
record_feedback |
write | Score a SME's session performance; drives the smoothed quality score |
propose_promotion |
write | Nominate a workspace SME for the shared library (auto-gated + moderated) |
review_promotion |
admin | Approve/reject the promotion queue |
Setup
1. Supabase
Create a project, then apply the migration:
supabase link --project-ref <ref>
supabase db push # applies supabase/migrations/*.sql
supabase functions deploy embed # optional: enables semantic/hybrid search
The embed edge function runs the built-in gte-small model. It's optional — without it,
search falls back to keyword FTS automatically.
2. Get an API key
Self-service (recommended): visit /dashboard, create an account (Supabase Auth), and
mint a key. First login auto-creates your workspace. Requires SUPABASE_ANON_KEY to be set and
email auth enabled in the Supabase project.
Operator / CLI: for admin keys or scripted provisioning:
SUPABASE_URL=... SUPABASE_SERVICE_ROLE_KEY=... \
node scripts/create-key.mjs --workspace "My Team" # read,write key
node scripts/create-key.mjs --workspace "Ops" --admin # admin key
Either way the plaintext key (sme_live_...) is shown once; only its SHA-256 hash is stored.
3. Deploy to Vercel
vercel deploy
Set env vars in the Vercel project: SUPABASE_URL, SUPABASE_SERVICE_ROLE_KEY,
SUPABASE_ANON_KEY (powers the dashboard), ANTHROPIC_API_KEY, and CRON_SECRET (protects
the daily /api/cron/maintenance job that recomputes quality scores, prunes rate-limit
windows, and backfills embeddings). Also set NEXT_PUBLIC_BASE_URL to your canonical site URL
and ADMIN_EMAILS (comma-separated) to bootstrap admin access.
3b. Email (confirmations & password resets)
Auth emails are sent by Supabase, not this app. If new accounts don't receive validation
emails, it's a Supabase config issue — most often that custom SMTP isn't configured (the
built-in sender is rate-limited to a few per hour, for testing only). In the Supabase dashboard:
configure Authentication → Emails → SMTP Settings (Resend/Postmark/SendGrid/SES), enable
Confirm email under Providers → Email, and add your site + …/auth/callback to the
URL Configuration allow-list.
4. Connect a client
claude mcp add --transport http sme-library \
https://<deployment>/api/mcp \
--header "Authorization: Bearer sme_live_..."
Works with Claude Code, claude.ai custom connectors, and any Streamable-HTTP MCP client.
Local development
cp .env.example .env # fill in Supabase + Anthropic
npm install
npm run dev # HTTP server on :3000 (endpoint: /api/mcp)
SME_API_KEY=sme_live_... npm run stdio # or stdio transport for local clients
Security model
- API keys are hashed at rest (SHA-256) and passed only in the
Authorizationheader — the credential never appears in tool arguments or the model's context. - Every key maps to one workspace with scopes (
read,write,promote,admin) and a rate-limit tier; per-key fixed-window limits are enforced in Postgres. - Supabase RLS is enabled deny-all on every table as defense-in-depth; only the server's service-role client can read or write.
- All writes land in
audit_log.
推荐服务器
Baidu Map
百度地图核心API现已全面兼容MCP协议,是国内首家兼容MCP协议的地图服务商。
Playwright MCP Server
一个模型上下文协议服务器,它使大型语言模型能够通过结构化的可访问性快照与网页进行交互,而无需视觉模型或屏幕截图。
Magic Component Platform (MCP)
一个由人工智能驱动的工具,可以从自然语言描述生成现代化的用户界面组件,并与流行的集成开发环境(IDE)集成,从而简化用户界面开发流程。
Audiense Insights MCP Server
通过模型上下文协议启用与 Audiense Insights 账户的交互,从而促进营销洞察和受众数据的提取和分析,包括人口统计信息、行为和影响者互动。
VeyraX
一个单一的 MCP 工具,连接你所有喜爱的工具:Gmail、日历以及其他 40 多个工具。
graphlit-mcp-server
模型上下文协议 (MCP) 服务器实现了 MCP 客户端与 Graphlit 服务之间的集成。 除了网络爬取之外,还可以将任何内容(从 Slack 到 Gmail 再到播客订阅源)导入到 Graphlit 项目中,然后从 MCP 客户端检索相关内容。
Kagi MCP Server
一个 MCP 服务器,集成了 Kagi 搜索功能和 Claude AI,使 Claude 能够在回答需要最新信息的问题时执行实时网络搜索。
e2b-mcp-server
使用 MCP 通过 e2b 运行代码。
Neon MCP Server
用于与 Neon 管理 API 和数据库交互的 MCP 服务器
Exa MCP Server
模型上下文协议(MCP)服务器允许像 Claude 这样的 AI 助手使用 Exa AI 搜索 API 进行网络搜索。这种设置允许 AI 模型以安全和受控的方式获取实时的网络信息。