@markygab/nest-mcp
NestJS package that provides decorator-based tool discovery, TypeBox/AJV validation, and MCP SDK server construction for building MCP servers.
README
@markygab/nest-mcp
Internal NestJS primitives for defining and serving Model Context Protocol (MCP) tools. The package provides decorator-based tool discovery, TypeBox/AJV input and output validation, explicit parameter injection, and MCP SDK server construction.
It deliberately does not include HTTP routing, authentication, authorization, tenant policy, or application-specific context. Each host application supplies those concerns at its own boundary.
Install and register
Add the workspace package as a dependency, then import NestMcpModule once in
the application's root module or a module visible to all MCP tool providers.
import { Module } from "@nestjs/common";
import { NestMcpModule } from "@markygab/nest-mcp";
import { ProjectsModule } from "./projects/projects.module.js";
@Module({
imports: [NestMcpModule, ProjectsModule],
})
export class AppModule {}
The package uses Nest's DiscoveryService to find decorated providers across
the application container. Tool classes must still be registered as normal
Nest providers in their feature modules.
Define a tool
Use TypeBox schemas for inputs and, where practical, outputs. @McpTools()
assigns the provider to one named MCP server. @McpTool() declares a single
tool, while @McpArgs() and @McpContext() make handler injection explicit.
import { Inject, Injectable } from "@nestjs/common";
import {
McpArgs,
McpContext,
McpTool,
McpTools,
type McpRequestContext,
type McpToolArgs,
} from "@markygab/nest-mcp";
import { Type } from "@sinclair/typebox";
import { ProjectsQueryService } from "./projects-query.service.js";
const GetProjectInput = Type.Object({
id: Type.String({ minLength: 1 }),
});
const GetProjectOutput = Type.Object({
project: Type.Union([
Type.Object({ id: Type.String(), name: Type.String() }),
Type.Null(),
]),
});
type GetProjectInput = McpToolArgs<typeof GetProjectInput>;
@Injectable()
@McpTools("projects_mcp")
export class ProjectsMcp {
constructor(
@Inject(ProjectsQueryService)
private readonly projects: ProjectsQueryService,
) {}
@McpTool({
name: "projects_get",
title: "Get project",
description: "Get a project visible to the current caller.",
inputSchema: GetProjectInput,
outputSchema: GetProjectOutput,
})
async getProject(
@McpArgs() input: GetProjectInput,
@McpContext() context: McpRequestContext,
) {
return {
project: await this.projects.getVisibleTo(context.actorId, input.id),
};
}
}
Tool names must be unique within a named server. Keep handlers thin: enforce application policy and call existing domain services rather than placing vendor-client or persistence logic in MCP classes.
Tool schemas
@McpTool() accepts TypeBox schemas and ordinary JSON Schema for both input
and output. TypeBox remains a convenient authoring option and
McpToolArgs<typeof schema> continues to infer its handler type. For plain
JSON Schema—such as a schema derived from an OpenAPI document—supply your own
generated or application type for the handler argument; arbitrary JSON Schema
is not inferred as TypeScript.
const CreateIssueInputSchema = {
type: "object",
properties: {
priority: { type: "string", enum: ["low", "high"] },
title: { type: "string" },
},
required: ["priority", "title"],
} as const;
type CreateIssueInput = {
priority: "low" | "high";
title: string;
};
@McpTool({
name: "issues_create",
description: "Create an issue",
inputSchema: CreateIssueInputSchema,
})
createIssue(@McpArgs() input: CreateIssueInput) {
return this.issues.create(input);
}
Guard tool invocations
@UseMcpGuards() attaches policy-agnostic guards to a tools class or an
individual tool. Guards run after the input schema is validated and before the
handler. Class guards run first, followed by method guards. Return false or
throw to prevent invocation; the caller receives the normal MCP tool error
response. Pass either a guard instance or an injectable Nest provider class;
provider classes are resolved from the application container.
import { Injectable } from "@nestjs/common";
import {
McpArgs,
McpTool,
McpTools,
UseMcpGuards,
type McpExecutionContext,
type McpGuard,
} from "@markygab/nest-mcp";
const auditGuard: McpGuard = {
canActivate(context: McpExecutionContext) {
// validatedArgs, serverName, toolName, requestContext, and toolOptions
// are available here. Application policy remains outside this library.
return true;
},
};
@Injectable()
class ProjectsAccessGuard implements McpGuard {
canActivate(context: McpExecutionContext) {
return Boolean(context.requestContext.actorId);
}
}
@McpTools("projects_mcp")
@UseMcpGuards(auditGuard, ProjectsAccessGuard)
class ProjectsMcp {
@McpTool({ name: "projects_get", description: "Get a project", inputSchema: GetProjectInput })
@UseMcpGuards({ canActivate: () => true })
getProject(@McpArgs() input: GetProjectInput) {
return { project: input.id };
}
}
Intercept tool invocations
@UseMcpInterceptors() wraps a permitted tool invocation. It accepts
McpInterceptor objects or injectable Nest provider classes and may be applied
to a tools class or a single tool. Provider classes are resolved from the
application container.
Interceptors receive the same McpExecutionContext as guards, including the
validated arguments and trusted request context. Class interceptors enter in
declaration order, then method interceptors; they unwind in reverse order.
Thrown errors propagate to the standard MCP tool error response.
import { Injectable } from "@nestjs/common";
import {
UseMcpInterceptors,
type McpExecutionContext,
type McpInterceptor,
} from "@markygab/nest-mcp";
@Injectable()
class TimingInterceptor implements McpInterceptor {
async intercept(context: McpExecutionContext, next) {
const startedAt = Date.now();
try {
return await next();
} finally {
// Hosts decide whether and where to record this information.
console.log(context.toolName, Date.now() - startedAt);
}
}
}
@UseMcpInterceptors(TimingInterceptor)
class ProjectsMcp {
// @UseMcpInterceptors(...) can also be applied to an @McpTool() method.
}
Request context and policy
McpRequestContext contains common request identifiers, optional actor and
tenant fields, an optional integration object, and allowedToolNames.
Applications can add their own fields when creating the context.
When a request context contains allowedToolNames, the package hides every
other tool from both tools/list and tools/call. Authentication and
authorization should run before the MCP request reaches the package; build the
trusted context from the verified principal and route parameters, never from
model-provided tool arguments.
Serve over HTTP
The host owns the controller, authentication, and request-context mapping. Add
a controller and transport service to the application that imports
NestMcpModule; the example below exposes POST, GET, and DELETE requests
at /mcp/projects_mcp, as required by Streamable HTTP clients.
import {
All,
Controller,
Inject,
Injectable,
Module,
Param,
Req,
Res,
} from "@nestjs/common";
import { StreamableHTTPServerTransport } from "@modelcontextprotocol/sdk/server/streamableHttp.js";
import { McpServerService, type McpRequestContext } from "@markygab/nest-mcp";
import type { Request, Response } from "express";
@Injectable()
export class ProjectsMcpHttpService {
constructor(
@Inject(McpServerService)
private readonly mcpServer: McpServerService,
) {}
async handle(
serverName: string,
request: Request,
response: Response,
context: McpRequestContext,
) {
const transport = new StreamableHTTPServerTransport({
// Use stateless requests when the host does not persist MCP sessions.
sessionIdGenerator: undefined,
});
const server = this.mcpServer.createServer(serverName, context);
await server.connect(transport);
try {
await transport.handleRequest(request, response, request.body);
} finally {
await transport.close();
}
}
}
@Controller("mcp")
export class ProjectsMcpHttpController {
constructor(
@Inject(ProjectsMcpAuthService)
private readonly auth: ProjectsMcpAuthService,
@Inject(ProjectsMcpHttpService)
private readonly http: ProjectsMcpHttpService,
) {}
@All(":serverName")
async handle(
@Param("serverName") serverName: string,
@Req() request: Request,
@Res() response: Response,
) {
const context = this.auth.verifyAuthorization(
request.header("authorization"),
serverName,
);
await this.http.handle(serverName, request, response, context);
}
}
Register these host-side providers alongside the shared module:
@Module({
imports: [NestMcpModule],
controllers: [ProjectsMcpHttpController],
providers: [ProjectsMcpAuthService, ProjectsMcpHttpService],
})
export class ProjectsMcpModule {}
ProjectsMcpAuthService should verify the bearer token and return only trusted
context, for example { requestId, actorId, tenantId, allowedToolNames }.
Bind the token to serverName so a credential issued for projects_mcp cannot
be reused against another named server. If you use stateful MCP sessions,
replace the stateless transport configuration with a host-managed session ID
strategy and retain the corresponding transport for later requests.
The host application owns the corresponding controller, authentication service, and transport service.
Runtime limits
Set these optional environment variables in the host process to bound tool execution and serialized responses:
NEST_MCP_TOOL_TIMEOUT_MSdefaults to15000.NEST_MCP_TOOL_RESPONSE_MAX_BYTESdefaults to128000.
Both values must be positive integers.
Serve over stdio
For a local process, retrieve McpServerService from the Nest application
context and provide the named server explicitly:
const app = await NestFactory.createApplicationContext(AppModule);
await app.get(McpServerService).startStdio("projects_mcp");
Exports
NestMcpModule- registers discovery, validation, invocation, and server providers.McpTools,McpTool,McpArgs,McpContext- tool-definition decorators.McpServerService- creates MCP SDK server instances, lists tools, and calls tools programmatically.McpRequestContext,McpToolArgs,McpToolOptions, and tool metadata types.
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。