AI导哈

MCP服务器交互记录回放工具

mcp-replay 是一个用于记录和回放MCP服务器交互的工具，旨在为CI测试提供确定性环境，类似于nock/msw但专为MCP设计。

  • by mstuart
  • TypeScript
  • 本地部署
  • 更新于
在 GitHub 查看 mstuart/mcp-replay

使用说明

Record and replay MCP server interactions for deterministic CI testing — nock/msw for MCP

Record and replay MCP (Model Context Protocol) server interactions for deterministic CI testing -- the nock/msw equivalent for MCP.

MCP is a JSON-RPC 2.0 protocol that lets AI models call tools and read resources from external servers.

The problem

Testing code that calls MCP servers is painful:

  • Live servers are slow, flaky, and cost money
  • Mocking JSON-RPC by hand is tedious and error-prone
  • Tests that depend on external services break in CI

mcp-replay lets you record real MCP responses to fixture files, then replay them in tests without a live server. Deterministic, fast, zero network calls.

Install

npm install mcp-replay --save-dev

Zero runtime dependencies -- uses only Node.js built-in crypto and fs.

Quick start

import { McpReplay } from 'mcp-replay';
import { describe, it, before, after } from 'node:test';
import assert from 'node:assert/strict';

describe('weather tool', () => {
  const replay = new McpReplay({
    fixtureDir: './fixtures',
    fallback: 'fail-fast',
  });

  before(async () => {
    // Load fixtures from disk
    await replay.start();

    // Or add fixtures manually
    replay.addFixture('tools/call', { name: 'get_weather', arguments: { city: 'London' } }, {
      jsonrpc: '2.0',
      id: 1,
      result: { content: [{ type: 'text', text: '15C, partly cloudy' }] },
    });
  });

  after(async () => {
    // Save any new fixtures to disk
    await replay.stop();
  });

  it('returns weather data', async () => {
    const client = replay.getClient();
    const result = await client.callTool('get_weather', { city: 'London' });
    assert.equal(result.content[0].text, '15C, partly cloudy');
  });
});

McpReplayClient

replay.getClient() returns a lightweight client that routes calls through the replay instance:

const client = replay.getClient();

// Call a tool
const result = await client.callTool('get_weather', { city: 'London' });
console.log(result.content[0].text); // "15C, partly cloudy"

// List tools
const { tools } = await client.listTools();
console.log(tools.map(t => t.name));

Recording fixtures

Option 1: Add fixtures in test setup

replay.addFixture('tools/call', { name: 'my_tool', arguments: { key: 'value' } }, {
  jsonrpc: '2.0',
  id: 1,
  result: { content: [{ type: 'text', text: 'response data' }] },
});

Option 2: Pre-populate fixture files

Create JSON files in your fixture directory:

fixtures/
  tools_call/
    <sha256-hash>.json
  tools_list/
    <sha256-hash>.json

Each fixture file:

{
  "method": "tools/call",
  "paramsHash": "abc123...",
  "params": { "name": "get_weather", "arguments": { "city": "London" } },
  "response": {
    "jsonrpc": "2.0",
    "id": 1,
    "result": { "content": [{ "type": "text", "text": "15C, partly cloudy" }] }
  },
  "recordedAt": "2025-01-15T10:30:00.000Z"
}

The hash is computed from SHA-256(JSON.stringify({ method, params })). You can use the exported hashRequest function to compute it:

import { hashRequest } from 'mcp-replay';
const hash = hashRequest('tools/call', { name: 'get_weather', arguments: { city: 'London' } });

Scrubbing secrets

Use scrubFields to strip sensitive fields from params before hashing and before saving to fixtures. This ensures that different API keys produce the same fixture match:

const replay = new McpReplay({
  fixtureDir: './fixtures',
  scrubFields: ['Authorization', 'apiKey', 'token'],
});

// These two calls will match the same fixture:
// { name: 'api', Authorization: 'Bearer abc', arguments: { q: 'test' } }
// { name: 'api', Authorization: 'Bearer xyz', arguments: { q: 'test' } }

Fields are removed recursively from all nested objects and arrays.

Options

Option Type Default Description
fixtureDir string (required) Directory for fixture files
strategy 'exact' | 'match-by-hash' 'match-by-hash' Matching strategy
fallback 'fail-fast' | 'pass-through' 'fail-fast' What to do on cache miss
scrubFields string[] [] Field names to remove before hashing

API

McpReplay

  • new McpReplay(opts: ReplayOptions) -- create an instance
  • start(): Promise<void> -- load fixtures from disk
  • stop(): Promise<void> -- save fixtures to disk
  • intercept(request: McpRequest): Promise<McpResponse> -- match a request to a fixture
  • addFixture(method, params, response): void -- add a fixture programmatically
  • getClient(): McpReplayClient -- get a client that routes through this instance

FixtureStore

  • get(method, params): Fixture | undefined
  • set(method, params, response): void
  • save(path): Promise<void>
  • load(path): Promise<void>
  • clear(): void

hashRequest(method, params?, scrubFields?): string

Compute the SHA-256 hash used for fixture matching.

scrub(obj, fields): unknown

Recursively remove fields from an object.

Project docs

License

MIT. See LICENSE.

相关 MCP