MCP Apps Example Gallery (fable candidate)
Hosts six official Model Context Protocol MCP Apps examples as remote Streamable HTTP MCP servers you can try by URL without cloning or local setup.
README
MCP Apps Example Gallery (fable candidate)
Six official Model Context Protocol MCP Apps examples,
hosted as remote Streamable HTTP MCP servers you can try by URL — no cloning,
no npx, no tunnel. Each example has a stable endpoint, a gallery card with a
screenshot and sample prompt, and a Copy MCP URL action.
Independence notice. This gallery is an independent hosted adaptation of the official examples, pinned at upstream commit
10195ad9. It is not an official Model Context Protocol service and is not hosted or endorsed by the Model Context Protocol project. It is a demonstration service: no accounts, no stored user data, and no SLA.
The apps
| Slug | Example | Demonstrates | MCP endpoint path |
|---|---|---|---|
get-time |
basic-server-react | Smallest tool + ui:// resource + UI-to-tool round trip |
/apps/get-time/mcp |
budget-allocator |
budget-allocator-server | Editable form, charts, recalculation | /apps/budget-allocator/mcp |
cohort-heatmap |
cohort-heatmap-server | Dense interactive data visualization | /apps/cohort-heatmap/mcp |
customer-segmentation |
customer-segmentation-server | Filtering and chart interaction | /apps/customer-segmentation/mcp |
scenario-modeler |
scenario-modeler-server | Scenario templates + custom 12-month projections | /apps/scenario-modeler/mcp |
transcript |
transcript-server | Live browser speech transcription (Web Speech API) | /apps/transcript/mcp |
Non-MCP routes: / (gallery page), /apps.json (machine-readable manifest),
/healthz, /readyz, /version.
Production endpoints
The production gallery is served at
https://openwork-mcp-app-gallery-fable.vercel.app (verified against the
exact deployed commit; receipts in BENCHMARK_REPORT.md and TIMELINE.md):
- https://openwork-mcp-app-gallery-fable.vercel.app/apps/get-time/mcp
- https://openwork-mcp-app-gallery-fable.vercel.app/apps/budget-allocator/mcp
- https://openwork-mcp-app-gallery-fable.vercel.app/apps/cohort-heatmap/mcp
- https://openwork-mcp-app-gallery-fable.vercel.app/apps/customer-segmentation/mcp
- https://openwork-mcp-app-gallery-fable.vercel.app/apps/scenario-modeler/mcp
- https://openwork-mcp-app-gallery-fable.vercel.app/apps/transcript/mcp
Try an example in under five minutes
- Open the gallery page and press Copy MCP URL on a card (or copy an endpoint above).
- Add the URL as a remote MCP server in an MCP Apps-compatible host. In OpenWork, add it as a user-configured MCP server; other hosts have an equivalent "add remote MCP server" flow.
- Send the card's sample prompt (for example, for
budget-allocator: "Create a $1 million seed-stage budget I can adjust interactively."). - Interact with the app UI that renders inside the conversation.
Hosts without MCP Apps support still work: every tool returns an ordinary text and/or structured result as fallback, so a plain MCP client gets useful output without any UI.
Protocol support: the current 2026-07-28 protocol revision (per-request
envelope) and the stateless 2025-era Streamable HTTP flow
(initialize → tools/* / resources/*) are both served on every endpoint.
GET without a session and DELETE return 405 — serving is stateless and
there is no server-side session to terminate.
Run it locally
Requires Node.js 24.x and pnpm 10.28.0 (via corepack):
corepack enable
pnpm install --frozen-lockfile
pnpm dev
This builds the six app UIs into the immutable resource bundle, generates the
gallery site, and serves everything at http://localhost:3000.
Reproduce the complete release gate (formatting, lint, strict typecheck, provenance/notices verification, source-boundary scan, dependency audit, 95-test unit/gateway/contract matrix over both protocol eras, the production Vercel build, the architecture invariant, SBOM generation, and 15 browser tests through the upstream basic host):
pnpm exec playwright install chromium
pnpm release:check
Source attribution and provenance
- Upstream example source is copied at the exact pinned commit under
upstream/ext-apps/, with per-file digests and every local modification documented inupstream/manifest.jsonand verified bypnpm verify:notices. - Local modifications are limited to: import adaptation from MCP SDK v1 +
@modelcontextprotocol/ext-apps/serverto the gallery's SDK v2 adapter (src/mcp-app-adapter.ts), reading UI HTML from the bundled immutable store instead of the local filesystem, and a deterministic seeded dataset forcustomer-segmentation(stateless serverless instances must serve identical data). - Gallery-owned code is Apache-2.0 (
LICENSE); copied upstream code keeps its own notices (THIRD_PARTY_NOTICES.md). pnpm check:upstreamreports upstream drift read-only; upstream is never fetched at build or runtime.
Data behavior, logging, and limits
- No accounts, no cookies, no database, no uploads, no stored user state.
All app data is synthetic or request-bounded; the
transcriptapp's speech recognition runs in your browser (the host asks for microphone permission) and no audio or transcript ever reaches the gallery server. - No intended server-side network egress in any Wave 1 app; CI scans the runtime source for network, subprocess, and filesystem-write primitives.
- Logs contain only anonymous aggregate metadata: app slug, JSON-RPC method category, status, duration, and byte counts — never tool arguments, results, prompts, resource contents, headers, cookies, or IP addresses.
- Limits: 256 KiB request bodies; 512 KiB tool results; 1 MiB bundled UI
resources; a 15-second application deadline (30-second platform maximum);
per-instance concurrency caps with
429+Retry-Aftershedding; platform edge rate limiting on/apps/*/mcp. - Apps can be disabled individually (fail-closed
DISABLED_APP_SLUGSoverride); disabled apps stay visible asenabled: falsein/apps.json.
No SLA
This is a hosted learning and demonstration surface, run on provider infrastructure without an availability commitment. Endpoints may be rate-limited, disabled, or removed at any time.
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。