[BUG] Claude Code SDK: SDK MCP server faills to connect due to closed stream

I encountered the same issue today and was able to work around it with a minimal adjustment:

1. Refactored into an MCP Server using @modelcontextprotocol/sdk (this was straightforward since the code format is nearly identical).
2. Updated the mcpServers definition as shown below.

Before:

for await (const message of query({
    prompt: ...,
    options: {
        allowedTools: [
            ...
            'mcp__test',
        ],
        mcpServers: {
            'test': testMcpServer,
        },
    },
})) {

After:

for await (const message of query({
    prompt: ...,
    options: {
        allowedTools: [
            ...
            'mcp__test',
        ],
        mcpServers: {
            'test': {
                type: 'stdio',
                command: 'npx',
                args: ['-y', 'tsx', 'src/mcps/test.ts'],
            },
        },
    },
})) {

With this approach, I was able to get things running in under five minutes with only minimal changes.
Depending on your setup, you may need to use a built JS artifact instead of tsx, but for my use case this was sufficient.

@selenehyun I'm glad to hear this works for you, but this does not run the local MCP server in the same process. That's my main motivation for using the SDK type.

This still reproduces with v1.0.96...

same issue for me

Broken with v1.0.98 as well

same issue

same issue , even without the allowedTools prop

Trying the example here https://docs.anthropic.com/en/docs/claude-code/sdk/sdk-typescript#creating-custom-tools but it doesn't seem to work either:

[ERROR] MCP server "my-custom-tools" Failed to connect SDK MCP server: Error: Stream closed

FWIW, I have the SDK MCP server type working in a single Bun-based project using v1.0.98, but I can't get it working in any other project. I have no idea why it works yet...

@codylund by any other project, do you mean projects using node or deno?

@ShuangLiu1992 Just any other Bun-based projects in general. I don't know what's special about the project where the SDK MCP server is functional. It's an overall basic set up and is very similar to the example I added in my original post. It makes me suspect some kind of obscure race condition in the Claude Code SDK.

It seems to still be an issue in v1.0.105

Same issue here

Running into this bug as well!

Found a workaround

import { createSdkMcpServer, query, tool } from "@anthropic-ai/claude-code"
import { z } from "zod/v3"

const options = {
  name: "my-custom-tools",
  version: "1.0.0",
  tools: [
    tool(
      "calculate_compound_interest",
      "Calculate compound interest for an investment",
      {
        principal: z.number().describe("Initial investment amount"),
        rate: z.number().describe("Annual interest rate (as decimal, e.g., 0.05 for 5%)"),
        time: z.number().describe("Investment period in years"),
        n: z.number().default(12).describe("Compounding frequency per year"),
      },
      async args => {
        console.log("calculate_compound_interest was called")
        
        const amount = args.principal * (1 + args.rate / args.n) ** (args.n * args.time)
        const interest = amount - args.principal

        return {
          content: [
            {
              type: "text",
              text: `Final amount: $${amount.toFixed(2)}\nInterest earned: $${interest.toFixed(2)}`,
            },
          ],
        }
      },
    ),
  ],
}

const customServer = createSdkMcpServer(options)

function userMessage(content) {
  return {
    type: "user",
    message: { role: "user", content: content },
  }
}

async function* prompt() {
  yield userMessage("Calculate compound interest for $10,000 at 5% for 10 years using the calculate_compound_interest tool")
  await new Promise(res => setTimeout(res, 10000))
}

for await (const message of query({
  prompt: prompt(),
  options: {
    permissionMode: "bypassPermissions",
    mcpServers: {
      "my-custom-tools": customServer,
    },
    maxTurns: 3,
  },
})) {
  if (message.type === "result") {
    console.log(message.result)
  }
}

The key is keeping the input stream artificially open via the await new Promise(res => setTimeout(res, 10000)) sleep. It isn't ideal, but there is likely a better workaround that closes the input stream once claude-code returns the result.

Looks like this was fixed in the Python Claude Code SDK: https://github.com/anthropics/claude-code-sdk-python/pull/157

same issue

same issue

I encountered the same issue today and was able to work around it with a minimal adjustment: 1. Refactored into an MCP Server using @modelcontextprotocol/sdk (this was straightforward since the code format is nearly identical). 2. Updated the mcpServers definition as shown below. Before: for await (const message of query({ prompt: ..., options: { allowedTools: [ ... 'mcp__test', ], mcpServers: { 'test': testMcpServer, }, }, })) { After: for await (const message of query({ prompt: ..., options: { allowedTools: [ ... 'mcp__test', ], mcpServers: { 'test': { type: 'stdio', command: 'npx', args: ['-y', 'tsx', 'src/mcps/test.ts'], }, }, }, })) { With this approach, I was able to get things running in under five minutes with only minimal changes. Depending on your setup, you may need to use a built JS artifact instead of tsx, but for my use case this was sufficient.

work for me

await new Promise(res => setTimeout(res, 10000))

Works for me as a workaround. Thanks for sharing @tariqkb ❤️

Here is the same workaround but not waiting for the full timeout simply by checking if "results" are in messages. Also no need to have bypassPermissions and you can allow and disallow tools. Thanks @claude for the help.

describe('SDK MCP Tools - TypeScript Translation', () => {
  let executions: string[];

  beforeEach(() => {
    executions = [];
  });

  it('should execute SDK MCP tools when allowed', { timeout: 35000 }, async () => {
    // Create echo tool that tracks executions
    const echoTool = tool(
      'echo',
      'Echo back the input text',
      {
        text: z.string().describe('Text to echo')
      },
      async (args) => {
        executions.push('echo');
        return {
          content: [{ type: 'text', text: `Echo: ${args.text}` }]
        };
      }
    );

    // Create test server with echo tool
    const server = createSdkMcpServer({
      name: 'test',
      version: '1.0.0',
      tools: [echoTool]
    });

    // Shared messages array that both generator and consumer can access
    const messages: any[] = [];

    // Create async generator that monitors for result message
    async function* promptGenerator() {
      yield {
        type: 'user' as const,
        message: { 
          role: 'user' as const, 
          content: 'Call the mcp__test__echo tool with text "Hello World"'
        }
      } as SDKUserMessage;
      
      // Wait for result message
      const maxChecks = 300; // Max 30 seconds (300 * 100ms)
      for (let i = 0; i < maxChecks; i++) {
        await new Promise(resolve => setTimeout(resolve, 100));
        
        // Check if we received a result message
        if (messages.some(m => m.type === 'result')) {
          break;
        }
      }
    }

    // Consume messages and populate the shared array
    for await (const message of query({
      prompt: promptGenerator(),
      options: {
        mcpServers: { test: server },
        allowedTools: ['mcp__test__echo'],
        maxTurns: 3
      }
    })) {
      messages.push(message);
    }

    // Verify the tool function was actually executed
    expect(executions).toContain('echo');
    expect(executions.length).toBe(1);
  });
}

Hi folks! Fix is ready and should be out soon in ~v1.0.121. You'll no longer need the timeout, but you will still need to provide a AsyncIterable<SDKUserMessage> streaming prompt (not a string prompt)

Hi @dltn i am using the 1.0128 version and the createSdkMcpServer creates the mcp server but in the init message it is always on status "failed". i am using an AsyncIterable and claude code is running but can't locate the custom tool, used exactly the same example in the docs.
any thoughts?

const customServer = createSdkMcpServer({
      name: 'weather-tools',
      tools: [
        tool(
          'get_weather',
          'Get current weather for a location',
          {
            location: z.string().describe('City name or coordinates'),
            units: z
              .enum(['celsius', 'fahrenheit'])
              .default('celsius')
              .describe('Temperature units'),
          },
          async (args) => {
            // Call weather API
            const response = await fetch(
              `https://api.weather.com/v1/current?q=${args.location}&units=${args.units}`,
            );
            const data = await response.json();

            return {
              content: [
                {
                  type: 'text',
                  text: `Temperature: ${data.temp}°\nConditions: ${data.conditions}\nHumidity: ${data.humidity}%`,
                },
              ],
            };
          },
        ),
      ],
    });

    async function* generateMessages() {
      yield {
        type: 'user' as const,
        message: {
          role: 'user' as const,
          content: 'what is the weather in Tokyo ? use the weather tool',
        },
      } as SDKUserMessage;
    }
    for await (const message of query({
      prompt: generateMessages(),
      options: {
        mcpServers: {
          test: customServer,
        },
        allowedTools: ['mcp__weather-tools__get_weather'],
      },
    })) {
      console.log(JSON.stringify(message, null, 2));
    }

and in the init message i am always seeing this:

"mcp_servers": [..., { "name" : "test" , "status": "failed" }]

which dosn't let's claude code use the whether tool

@tom-n-terra What does your npm list claude-code say? With @anthropic-ai/claude-cli@2.0.3, your test file works for me:

  "mcp_servers": [
    {
      "name": "test",
      "status": "connected"
    }

Going to close this. Please file another issue if you experience this again!

This issue has been automatically locked since it was closed and has not had any activity for 7 days. If you're experiencing a similar issue, please file a new issue and reference this one if it's relevant.