callspec
Provides an MCP server that automatically exposes API routes as MCP tools with input validation and authentication, enabling AI agents to call the same handlers as HTTP RPC.
README
<div align="center"> <picture> <source srcset="assets/callspec-lockup-light.svg?cb=2" media="(prefers-color-scheme: light)" /> <source srcset="assets/callspec-lockup-dark.svg?cb=2" media="(prefers-color-scheme: dark)" /> <img src="assets/callspec-lockup-dark.svg?cb=2" alt="callspec" /> </picture>
<h3 align="center">One spec ships API, docs UI, OpenAPI, MCP, and typed clients.</h3>
<br>
<p> <a href="assets/callspec-ui-chirp-demo-home.png"> <img src="assets/callspec-ui-chirp-demo-home.png" alt="callspec UI" width="920" /> </a> </p> </div>
Define your API once and get an HTTP RPC server, white-label docs, OpenAPI 3.1, an MCP server, and a typed client. No duplicate schemas, no bolt-on doc stack, no hand-maintained tool manifests.
Every API and MCP call gets input validation at the boundary with clear error messages. TypeScript clients get compile-time type checking and LSP autocomplete from the same spec — with simple, clean, human-readable types.
| Feature | Location |
|---|---|
| HTTP RPC API | POST /v1/<methodName> |
| Interactive UI docs | /docs |
| OpenAPI 3.1 | /openapi.json |
| MCP tools | /mcp |
| Typed client | client<API['searchRecent']>('searchRecent', input) |
| Input validation | Runtime & compile-time (TypeScript) |
Complete Example
import express from 'express';
import {defineSpec, defineRoute, mountSpec} from 'callspec';
import {predicates as p} from 'runtyp';
type AuthContext = {userId: string};
async function searchRecent(
input: {query: string; max_results?: number},
ctx: AuthContext,
) {
return {
results: [{id: '1', text: `Match for "${input.query}"`, authorId: ctx.userId}],
count: 1,
};
}
const api = defineSpec({
searchRecent: defineRoute({
input: p.object({
query: p.string({description: 'Search query (supports from:, #hashtag, …)'}),
max_results: p.optional(p.number({range: {min: 1, max: 100}})),
}),
meta: {
summary: 'Search recent posts',
tags: ['posts'],
},
access: 'private',
mcp: true,
handler: searchRecent,
}),
});
const app = express();
const router = express.Router();
router.use(express.json());
mountSpec(router, api, {
contextResolver: (req) => {
if (req.headers.authorization?.startsWith('Bearer ')) {
return {userId: 'user_123'};
}
return undefined;
},
docs: {
openApi: {title: 'My API', version: '1.0.0'},
ui: {
branding: {name: 'My API'},
},
},
});
app.use('/v1', router);
const port = Number(process.env.PORT ?? 3000);
app.listen(port, () => {
console.log(`RPC: http://127.0.0.1:${port}/v1/searchRecent`);
console.log(`Docs: http://127.0.0.1:${port}/v1/docs`);
console.log(`OpenAPI: http://127.0.0.1:${port}/v1/openapi.json`);
console.log(`MCP: http://127.0.0.1:${port}/v1/mcp`);
console.log('Auth: Authorization: Bearer anything (demo token)');
});
Getting started
npm i callspec runtyp express
npm i -D tsx typescript @types/express
Requirements: Node.js 18+, TypeScript 5+, Express 4.x (peer).
Try the demo (in this repo):
npm run build && npm run dev:docs
Open http://127.0.0.1:3456/v1/docs — Chirp sample API. Use Authorization: Bearer demo for private routes and MCP tools.
🔌 Built-in MCP server
Set mcp: true on any defineRoute. When any route opts in, mountSpec mounts MCP at /mcp automatically (override with mcp: { path, serverInfo, instructions }, or mcp: false to disable).
Agents call the same handlers as HTTP RPC — same auth gate, same input validation. Public tools work without a token; private tools return 401 without one. Every tools/call runs through the same runtyp pipeline: field { description }, ranges, and enums flow into MCP inputSchema; invalid args return structured validation errors (not opaque 500s) that agents can read and retry.
📖 callspec UI
Minimal, fast docs UI baked right into the package. Browse routes, try RPCs, read OpenAPI, and connect MCP clients from the home page. Point docs.ui.branding at your product — display name, welcome copy, website link, and logo (light/dark, optional brandAssetsDir for static files at /docs/brand/). Run npm run dev:docs to see the Chirp sample — callspec UI white-labeled as a fictional API.
Env-gated in production; flip on with:
docs: {
openApi: {title: 'Chirp API v2', version: '2.0.0'},
exposeOpenApi: true,
exposeUi: true,
openApiPath: '/openapi.json',
uiPath: '/docs',
ui: {
branding: {
name: 'Chirp',
intro: 'Read and write posts, timelines, lists, and direct messages.',
websiteUrl: 'https://chirp.social',
websiteLabel: 'chirp.social',
logoUrl: './brand/mark.png',
logoUrlDark: './brand/mark.png',
logoSize: 80,
},
brandAssetsDir: '/path/to/your/logos',
mcp: {authHint: 'Use Authorization: Bearer demo for private tools in this demo.'},
},
}
Toggle each surface independently — OpenAPI only, UI only, MCP off (mcp: false), or neither (docs: false).
Light and dark lockups follow prefers-color-scheme in docs; the UI footer switches marks on data-theme the same way as Castellan.
🔐 Auth
access: 'public'— no credentials required (e.g. Chirphealthcheck,getTweet)access: 'private'(default) — 401 withoutcontextResolverresult- App-specific auth stays in your
contextResolver(e.g. mapAuthorization: Bearer …to{userId, username})
Private gate runs before validation so unauthenticated callers never see field-level errors.
🧩 runtyp + OpenAPI
Field { description } on runtyp preds flows to JSON Schema → OpenAPI → callspec UI → MCP inputSchema. Route-level meta (summary, tags) is callspec-only.
Powered by runtyp for validation and schema generation.
📦 Client
Fetch-only — works in the browser and in Node 18+ (global fetch). The callspec/client entry has no http, https, or Express imports, so it is safe in frontend bundles.
Browser or frontend bundler — import the client subpath so you do not pull server code:
import type {API} from './my-api';
import {client} from 'callspec/client';
const results = await client<API['searchRecent']>('searchRecent', {
query: 'callspec',
max_results: 10,
}, {
endpoint: 'https://api.example.com/v1',
fetchOptions: {headers: {Authorization: `Bearer ${token}`}},
});
Service package — export the API type once from your spec:
import type {InferSpec} from 'callspec';
export const api = defineSpec({
getTweet: defineRoute({ /* … */ }),
searchRecent: defineRoute({ /* … */ }),
createTweet: defineRoute({ /* … */ }),
});
export type API = InferSpec<typeof api>;
Responses deserialize ISO date strings back to Date on read (deserializeResponse).
🛠 Development
npm run validate # build server + callspec UI, lint, test (incl. integration)
npm run dev:docs # Chirp demo API + callspec UI at :3456/v1/docs
Integration tests spin up Express in-process and verify OpenAPI, /docs, auth, MCP, and RPC end-to-end.
🤝 Help build the standard
callspec is early — and we're looking for maintainers and contributors who want to help define how typed APIs work in the age of agents.
The goal is simple: one spec → HTTP RPC, docs, OpenAPI, and MCP — no duplicate schemas, no bolt-on tool manifests, no duct-tape between surfaces.
If you join now, you're not polishing someone else's finished spec. You're shaping the defaults: callspec UI UX, MCP ergonomics, client DX, framework adapters, examples, and the docs people copy from. Early contributors tend to become the people others cite — show up in release notes, speak at the meetup, get asked "who built this?" when the pattern spreads.
Good first contributions: callspec UI polish, MCP client configs, docs and demos, runtyp/OpenAPI edge cases, Fastify/Hono mounts, issue triage, or a blog post about your integration.
- Issues & ideas: github.com/logfoxai/callspec/issues
- PRs welcome —
npm run validatebefore you push; conventional commits (feat:,fix:,docs:,style:, etc.)
If you want maintainer access or a dedicated area to own (callspec UI, MCP, clients, docs), open an issue or PR and say hi. We'd rather have a small crew that cares than a huge drive-by.
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。