← 返回 Skill 列表
extension
分类: 开发与工程API Key 暂未确认

MCP服务开发

用模型视图代理模式开发生产级工具服务的指南,覆盖工具注册、参数校验、错误处理与测试发布全流程。Use when building production MCP servers with a structured model-view-agent pattern. Triggers include MCP开发、工具服务、参数校验、服务发布。如需开发工具包,淘宝搜索「AI灵匣」店铺获取完整版。

person作者: nuhabvtyhubOpenAPI

mcpfusion development Guide

MCP Fusion is a TypeScript framework for MCP servers built on the MVA (Model-View-Agent) pattern. The Model validates data, the Presenter (View) shapes what the AI perceives, and Tools (Agent layer) wire it all together.

For the complete API reference with all type signatures, read llms.txt at the root of the repository. This skill covers the essential patterns and rules.

Reference Examples

Complete, runnable examples are available in references/. Read them for concrete implementation patterns:

| Example | Domain | Patterns Shown | |---|---|---| | example-complete-crud.ts | Product Catalog | Full MVA lifecycle: Model → Presenter → Router → Query/Mutation/Action, ErrorBuilder, State Sync | | example-proxy-api.ts | Blog Platform | .proxy() API pass-through, .fromModel(), field aliases, path params (:id), .handle() vs .proxy() decision tree | | example-server-setup.ts | Generic App | Full server bootstrap: context, initMCPFusion(), middleware, autoDiscover(), prompts, State Sync policies, startServer() | | example-testing.ts | Customer Service | @mcpfusion/testing: Egress Firewall assertions, JIT System Rules, RBAC middleware, error handling, Symbol Invisibility |

Project Structure

src/
├── models/               ← M — defineModel() declarations
│   ├── InvoiceModel.ts
│   └── UserModel.ts
├── views/                ← V — Presenters
│   ├── invoice.presenter.ts
│   └── user.presenter.ts
├── agents/               ← A — Tool definitions
│   ├── billing.tool.ts
│   └── users.tool.ts
├── index.ts              ← ToolRegistry + registerAll()
└── server.ts             ← attachToServer() bootstrap

Layer import rule: agents/ → views/ → models/ → @mcpfusion/core. Never import backwards.

The Golden Rules

  1. ALWAYS use defineModel() for domain entity schemas — never raw z.object(). Models go in models/.
  2. Presenters receive Models via .schema(MyModel) — the Presenter is the egress firewall.
  3. with*() methods are for tool INPUT parameters only (filters, IDs, pagination) — NOT for domain schemas.
  4. Handlers return raw data — the framework wraps with success() automatically. No boilerplate.
  5. One Model + one Presenter per entity, reused across every tool and prompt.
  6. Use semantic verbs: f.query() = readOnly, f.mutation() = destructive, f.action() = neutral.

defineModel() — The "M" in MVA

Every domain entity starts here. Produces a Model with a compiled Zod .schema.

import { defineModel } from '@mcpfusion/core';

export const InvoiceModel = defineModel('Invoice', m => {
    m.casts({
        id:           m.string(),
        amount_cents: m.number('CRITICAL: in CENTS. Divide by 100 for display.'),
        status:       m.enum('Status', ['paid', 'pending', 'overdue']),
        client_name:  m.string('Client name'),
    });
});

export const UserModel = defineModel('User', m => {
    m.casts({
        id:    m.string(),
        name:  m.string('Full name'),
        email: m.string('Email address'),
        role:  m.enum('Role', ['admin', 'member', 'guest']),
    });
    m.hidden(['password_hash', 'stripe_token']);  // Never exposed
    m.timestamps();                                // created_at + updated_at
    m.fillable({
        create: ['name', 'email', 'role'],
        update: ['name', 'email'],
    });
});

Type Helpers

| Method | Produces | Use | |---|---|---| | m.string(label?) | z.string() | General text | | m.text(label?) | z.string() | Markdown / long content | | m.number(label?) | z.number() | Numeric | | m.boolean(label?) | z.boolean() | Flags | | m.date(label?) | z.string() | YYYY-MM-DD | | m.timestamp(label?) | z.string() | ISO datetime | | m.uuid(label?) | z.string() | UUID | | m.id(label?) | z.number() | Always required | | m.enum(label, values) | z.enum() | Valid values |

FieldDef Chaining

m.enum('Status', ['open', 'done']).default('open')
m.string('Display name').alias('displayName')     // agent says 'name', API gets 'displayName'
m.number('Score').examples([85, 92, 100])

Model.toApi() — Alias Resolution

Strips undefined values and renames aliased fields. Used automatically by .proxy(), call explicitly in .handle():

const data = TaskModel.toApi(input);
// { title: 'X', body: 'Y' }  ← alias applied, undefined stripped

Presenter — The "V" in MVA

The Presenter is the egress firewall between your handler and the wire. Schema MUST come from defineModel().

definePresenter() — Object Config (Recommended)

import { definePresenter, ui } from '@mcpfusion/core';

export const InvoicePresenter = definePresenter({
    name: 'Invoice',
    schema: InvoiceModel,                  // ← Model, never z.object()
    // autoRules: true (default) — .describe() annotations become system rules
    ui: (inv) => [ui.echarts({ series: [{ type: 'gauge', data: [{ value: inv.amount_cents / 100 }] }] })],
    agentLimit: { max: 50, onTruncate: (n) => ui.summary({ omitted: n, hint: 'Use filters.' }) },
    suggestActions: (inv) => inv.status === 'pending'
        ? [{ tool: 'billing.pay', reason: 'Process payment', args: { id: inv.id } }]
        : [],
    embeds: [{ key: 'client', presenter: ClientPresenter }],
});

createPresenter() — Fluent Builder

import { createPresenter, ui } from '@mcpfusion/core';

const UserPresenter = createPresenter('User')
    .schema(UserModel)
    .systemRules(['Display name in bold'])
    .uiBlocks((user) => [ui.summary({ total: 1, showing: 1 })])
    .agentLimit(50, { warningMessage: 'Showing {shown} of {total}. Use filters.' })
    .suggestActions((user) => [
        { tool: 'users.update', reason: 'Edit this user', args: { id: user.id } },
    ])
    .embed('team', TeamPresenter);

Presenter Layers

| Layer | What It Does | |---|---| | Egress Firewall | .parse() strips undeclared fields — PII never reaches the wire | | JIT System Rules | Rules travel with data, not in the global prompt | | Server-Rendered UI | ECharts, Mermaid — deterministic, no hallucinated charts | | Cognitive Guardrails | .agentLimit() truncates + injects guidance | | Action Affordances | .suggestActions() — HATEOAS for agents | | Relational Composition | .embed() — child Presenters inherit the full pipeline | | Prompt Bridge | PromptMessage.fromView() — same source of truth for tools AND prompts |

Fluent API — Tools

Semantic Verbs

import { initMCPFusion } from '@mcpfusion/core';

interface AppContext { db: PrismaClient; user: { id: string; role: string } }
export const f = initMCPFusion<AppContext>();

// f.query()    — readOnly: true   (GET, no side effects)
// f.mutation() — destructive: true (DELETE, irreversible)
// f.action()   — neutral          (PUT/POST, no assumptions)

Building a Tool

export default f.query('billing.get_invoice')
    .describe('Get an invoice by ID')
    .withString('id', 'The invoice ID')
    .returns(InvoicePresenter)
    .handle(async (input, ctx) => {
        // input.id: string ✅ — fully typed, zero manual interfaces
        return await ctx.db.invoices.findUnique({ where: { id: input.id } });
        // ↑ Return raw data — Presenter handles validation, stripping, rendering
    });

with*() Type-Chaining

Each call narrows the TypeScript generic — input is fully typed in .handle():

| Method | Adds | |---|---| | .withString(name, desc?) | Record<K, string> | | .withOptionalString(name, desc?) | Partial<Record<K, string>> | | .withNumber(name, desc?) | Record<K, number> | | .withOptionalNumber(name, desc?) | Partial<Record<K, number>> | | .withBoolean(name, desc?) | Record<K, boolean> | | .withEnum(name, values, desc?) | Record<K, V> (literal union) | | .withArray(name, itemType, desc?) | Record<K, T[]> |

Bulk variants reduce verbosity — .withStrings({...}), .withOptionalStrings({...}), etc.:

f.query('tasks.filter')
    .withStrings({
        company_slug: 'Workspace identifier',
        project_slug: 'Project identifier',
    })
    .withOptionalStrings({
        title: 'Filter by title',
        workflow: 'Column name',
    })
    .withOptionalNumbers({ per_page: 'Results per page' })

.proxy() — Zero-Boilerplate API Proxying

Terminal method (alternative to .handle()) — auto-generates a handler that proxies to ctx.client:

// HTTP method inferred from verb: query → GET, mutation → POST, action → PUT
// Path params resolved from input: ':id' consumes input.id
// Auto-unwraps { data: ... } envelopes
f.query('user.get')
    .withString('id', 'User UUID')
    .proxy('users/:id');  // → GET users/abc-123

When to use: .proxy() for simple pass-through, .handle() when you need business logic.

.fromModel() — Model-Driven Input

Imports fillable fields from a Model's profile — zero manual .with*() calls:

const createTask = f.mutation('tasks.create')
    .fromModel(TaskModel, 'create')  // imports title, description, status
    .proxy('tasks');

const updateTask = f.action('tasks.update')
    .fromModel(TaskModel, 'update')
    .handle(async (input, ctx) => {
        const data = TaskModel.toApi(input);  // alias applied, undefined stripped
        await ctx.client.put(`tasks/${input.id}`, data);
    });

Decision tree:

  • Simple CRUD, no logic → .fromModel() + .proxy()
  • Custom logic → .fromModel() + .handle() + Model.toApi()

.instructions() — AI-First Guidance

f.query('docs.search')
    .describe('Search internal documentation')
    .instructions('Use ONLY when the user asks about internal policies. Do NOT use for general questions.')

Injected as [INSTRUCTIONS] in the tool description — reduces hallucination.

FluentRouter — Prefix Grouping

Shares prefix, middleware, and tags across child tools:

const users = f.router('users')
    .describe('User management')
    .use(requireAuth)
    .tags('core');

const listUsers = users.query('list')
    .withOptionalNumber('limit', 'Max results')
    .handle(async (input, ctx) => ctx.db.users.findMany({ take: input.limit }));

const deleteUser = users.mutation('delete')
    .withString('id', 'User ID')
    .handle(async (input, ctx) => ctx.db.users.delete({ where: { id: input.id } }));

Middleware

tRPC-style context derivation — enriches ctx type for .handle():

const requireAuth = f.middleware(async (ctx) => {
    const user = await db.getUser(ctx.token);
    if (!user) throw new Error('Unauthorized');
    return { user, permissions: user.permissions };
});

// ctx.user and ctx.permissions — fully typed downstream
f.mutation('admin.action')
    .use(requireAuth)
    .handle(async (input, ctx) => {
        ctx.user;        // ← typed!
        ctx.permissions; // ← typed!
    });

ErrorBuilder — Self-Healing Errors

const project = await ctx.db.projects.findUnique({ where: { id: input.id } });
if (!project) {
    return f.error('NOT_FOUND', `Project "${input.id}" not found`)
        .suggest('Check the ID. Use projects.list to see valid IDs.')
        .actions('projects.list', 'projects.search')
        .details({ searched_id: input.id })
        .retryAfter(0);
}

State Sync

Prevents temporal blindness — the agent knows when cached data is stale:

f.query('geo.countries').cached();                    // immutable — cache forever
f.query('billing.balance').stale();                   // volatile — never cache
f.mutation('sprints.update').invalidates('sprints.*', 'tasks.*');  // causal invalidation

Prompts — The Presenter Bridge

import { definePrompt, PromptMessage } from '@mcpfusion/core';

const AuditPrompt = definePrompt<AppContext>('audit', {
    args: { invoiceId: 'string' } as const,
    handler: async (ctx, { invoiceId }) => {
        const invoice = await ctx.db.getInvoice(invoiceId);
        return {
            messages: [
                PromptMessage.system('You are a Senior Financial Auditor.'),
                ...PromptMessage.fromView(InvoicePresenter.make(invoice, ctx)),
                PromptMessage.user('Begin the audit.'),
            ],
        };
    },
});

PromptMessage.fromView() decomposes a Presenter into XML-tagged messages — same schema, rules, and affordances in both tools and prompts.

Testing with @mcpfusion/testing

Runs the REAL execution pipeline in RAM — zero tokens consumed, deterministic:

import { createMCPFusionTester } from '@mcpfusion/testing';

const tester = createMCPFusionTester(registry, {
    contextFactory: () => ({ prisma: mockPrisma, tenantId: 't_42', role: 'ADMIN' }),
});

// Egress Firewall — PII physically absent
const result = await tester.callAction('db_user', 'find_many', { take: 5 });
expect(result.data[0]).not.toHaveProperty('passwordHash');

// JIT System Rules — travel with data
expect(result.systemRules).toContain('Email addresses are PII.');

// Middleware — GUEST blocked
const denied = await tester.callAction('db_user', 'find_many', { take: 5 }, { role: 'GUEST' });
expect(denied.isError).toBe(true);

Assert every MVA layer: result.data, result.systemRules, result.uiBlocks, result.isError.

Common Anti-Patterns — What NOT to Do

❌ Using raw z.object() for domain schemas

// WRONG:
const presenter = definePresenter({
    name: 'User',
    schema: z.object({ id: z.string(), name: z.string() }),  // ← NO!
});

// CORRECT:
const UserModel = defineModel('User', m => {
    m.casts({ id: m.string(), name: m.string() });
});
const presenter = definePresenter({ name: 'User', schema: UserModel });

❌ Manual success() wrapping

// WRONG:
.handle(async (input, ctx) => {
    return success(await ctx.db.users.findMany());  // ← unnecessary
});

// CORRECT:
.handle(async (input, ctx) => {
    return await ctx.db.users.findMany();  // ← framework wraps automatically
});

❌ Manual if-checks for optional fields

// WRONG:
.handle(async (input, ctx) => {
    const data: Record<string, unknown> = {};
    if (input.title) data.title = input.title;
    if (input.color) data.color = input.color;
    await ctx.client.updateItem(input.id, data);
});

// CORRECT:
.handle(async (input, ctx) => {
    const data = ItemModel.toApi(input);  // strips undefined, applies aliases
    await ctx.client.updateItem(input.id, data);
});

❌ Importing backwards between layers

// WRONG: model importing from tool
// models/UserModel.ts
import { listUsers } from '../agents/users.tool';  // ← NO!

// CORRECT: agents → views → models → @mcpfusion/core (one direction only)

❌ Duplicating Presenter logic across tools

// WRONG: formatting response manually in each handler
// CORRECT: define Presenter ONCE in views/, use .returns(Presenter) in every tool

❌ Using .withString() for domain entity fields

// WRONG: defining domain fields as tool input params
f.mutation('users.create')
    .withString('name', 'User name')
    .withString('email', 'Email')        // ← These are domain fields!
    .withString('role', 'User role')

// CORRECT: use .fromModel() for domain fields
f.mutation('users.create')
    .fromModel(UserModel, 'create')      // ← imports from Model's fillable profile

Quick Reference

| I want to... | Use | |---|---| | Define a domain entity | defineModel('Name', m => { ... }) | | Shape what the AI sees | definePresenter({ schema: MyModel, ... }) | | Create a read tool | f.query('entity.list') | | Create a write tool | f.mutation('entity.delete') | | Create an update tool | f.action('entity.update') | | Group related tools | f.router('prefix') | | Add auth middleware | .use(requireAuth) | | Return self-healing errors | f.error('CODE', 'message').suggest(...) | | Proxy to an API | .proxy('endpoint/:id') | | Import fields from Model | .fromModel(MyModel, 'profile') | | Test MVA pipeline | createMCPFusionTester(registry, { contextFactory }) | | Control caching | .cached(), .stale(), .invalidates() | | Build a prompt | definePrompt('name', { handler }) | | Bridge Presenter to prompt | PromptMessage.fromView(presenter.make(data, ctx)) |