mcp-use - Connect Any LLM to Any MCP Server

Code Execution Mode allows agents to interact with MCP tools by writing and executing JavaScript/TypeScript code instead of calling tools individually. This approach can significantly improve efficiency for complex tasks, data processing, and multi-step workflows. Based on Anthropic’s research on code execution with MCP, this feature enables agents to process data more efficiently and reduce context usage.

Code Execution Mode runs code in sandboxed environments. The default VM executor provides basic isolation, while the E2B executor provides strong cloud-based isolation. Choose the appropriate executor based on your security requirements.

Quick Start

Basic Setup (VM Executor)

The simplest way to enable Code Mode with local execution:

import { MCPClient } from "mcp-use";
import { MCPAgent } from "mcp-use/agent";
import { PROMPTS } from "mcp-use/client/prompts";

const client = new MCPClient(
  {
    mcpServers: {
      filesystem: {
        command: "npx",
        args: ["-y", "@modelcontextprotocol/server-filesystem", "./data"],
      },
    },
  },
  { codeMode: true } // Enable code mode with default VM executor
);

const agent = new MCPAgent({
  llm,
  client,
  systemPrompt: PROMPTS.CODE_MODE, // Use code mode instructions
});

// Agent can now write code to use tools
const result = await agent.run("List all files and count them");

E2B Setup (Cloud Sandbox)

For production or untrusted code, use E2B for stronger isolation:

const client = new MCPClient(
  {
    /* server config */
  },
  {
    codeMode: {
      enabled: true,
      executor: "e2b",
      executorOptions: {
        apiKey: process.env.E2B_API_KEY!,
        timeoutMs: 300000, // 5 minutes (default)
      },
    },
  }
);

VM with Custom Timeout

Customize the VM executor timeout:

const client = new MCPClient(
  {
    /* server config */
  },
  {
    codeMode: {
      enabled: true,
      executor: "vm", // Default, can be omitted
      executorOptions: {
        timeoutMs: 60000, // 1 minute (default: 30 seconds)
        memoryLimitMb: 512, // Optional memory limit
      },
    },
  }
);

Architecture

Code Mode uses a layered architecture to provide flexible code execution with MCP tool access:

┌────────────────────────────────────────────────────────────┐
│                       AI Agent                             │
│  (Writes JavaScript code to accomplish tasks)              │
└────────────────────────────┬───────────────────────────────┘
                             │
                             ▼
┌────────────────────────────────────────────────────────────┐
│                    CodeModeConnector                       │
│  Provides two special tools:                               │
│  • execute_code: Run JavaScript with tool access           │
│  • search_tools: Discover available tools                  │
└────────────────────────────┬───────────────────────────────┘
                             │
                    ┌────────┴────────┐
                    ▼                 ▼
        ┌───────────────────┐  ┌──────────────────┐
        │   Code Executor   │  │   MCP Servers    │
        │  (VM or E2B)      │  │  (github, etc.)  │
        │                   │  │                  │
        │  Runs code with   │  │  Provide tools   │
        │  tools exposed    │  │  for execution   │
        └───────────────────┘  └──────────────────┘

Flow

Agent discovers tools using search_tools
Agent writes code that calls tools as async functions
Code executes in VM or E2B sandbox
Tools are called either directly (VM) or via bridge (E2B)
Results are returned to the agent

Executor Types

Code Mode supports three executor types:

VM Executor (Default)

Local execution using Node.js vm module. Fast and simple, suitable for trusted environments. Advantages:

Zero latency - direct function calls
No external dependencies
Immediate execution
No cost

Limitations:

Basic isolation (not suitable for untrusted code)
Limited to Node.js built-ins
Shares host resources

Configuration:

const client = new MCPClient(config, {
  codeMode: true,
  // VM is the default, no need to specify codeExecutor
});

E2B Executor

Remote execution in E2B cloud sandboxes. Provides strong isolation for production use. Advantages:

True isolation in cloud sandbox
Full Linux environment
Network and filesystem access
Persistent sandboxes
Suitable for untrusted code

Limitations:

Network latency overhead
Requires E2B API key and billing
More complex debugging

Setup:

Install dependency:

yarn add @e2b/code-interpreter

Get API key: Visit e2b.dev and create an account to get your API key.
Configure client:

const client = new MCPClient(config, {
  codeMode: true,
  codeExecutor: "e2b",
  e2bApiKey: process.env.E2B_API_KEY,
  e2bTimeoutMs: 300000, // Optional: sandbox timeout (default 5 min)
});

Set environment variable:

export E2B_API_KEY=your_api_key_here

Custom Executor

Implement your own execution logic for specialized requirements. Custom Function:

const client = new MCPClient(config, {
  codeMode: {
    enabled: true,
    executor: async (code: string, timeout?: number) => {
      // Your custom execution logic
      const result = await myCustomRuntime.execute(code);

      return {
        result: result.value,
        logs: result.console,
        error: result.error,
        execution_time: result.duration,
      };
    },
  },
});

Custom Class:

import { BaseCodeExecutor } from "mcp-use";

class MyExecutor extends BaseCodeExecutor {
  async execute(code: string, timeout?: number): Promise<ExecutionResult> {
    await this.ensureServersConnected(); // Connect to MCP servers
    const namespaces = this.getToolNamespaces(); // Get available tools

    // Your execution logic here
    return { result, logs, error, execution_time };
  }

  async cleanup(): Promise<void> {
    // Clean up resources
  }
}

const client = new MCPClient(config, {
  codeMode: {
    enabled: true,
    executor: new MyExecutor(client),
  },
});

How Agent Code Works

When Code Mode is active, the agent writes JavaScript code that has access to:

1. Tool Namespaces

Tools are exposed as serverName.toolName(args) async functions:

// Call GitHub MCP server tools
const prs = await github.list_pull_requests({
  owner: "facebook",
  repo: "react",
});

// Call filesystem MCP server tools
const files = await filesystem.list_directory({ path: "/data" });

2. Tool Discovery

Use search_tools() to discover available tools at runtime:

// Find all tools (returns full schemas)
const allTools = await search_tools();

// Search for specific tools
const githubTools = await search_tools("github");

// Get just names and servers
const toolNames = await search_tools("", "names");

// Get names and descriptions
const toolDescs = await search_tools("", "descriptions");

3. Standard JavaScript

Full access to JavaScript built-ins for data processing:

// Filter and map data
const openBugs = prs.filter(
  (pr) => pr.labels.some((l) => l.name === "bug") && pr.state === "open"
);

// Aggregate results
const summary = {
  total: prs.length,
  open: openBugs.length,
  titles: openBugs.map((pr) => pr.title),
};

return summary;

Complete Example

Here’s how an agent might write code to analyze GitHub issues:

// 1. Discover what tools are available
const tools = await search_tools("github issue");
console.log(`Found ${tools.length} GitHub issue tools`);

// 2. Fetch issues from repository
const issues = await github.list_issues({
  owner: "modelcontextprotocol",
  repo: "servers",
  state: "open",
});

// 3. Process data efficiently in code
const criticalIssues = issues.filter((issue) =>
  issue.labels.some((label) => label.name === "critical")
);

const bugIssues = issues.filter((issue) =>
  issue.labels.some((label) => label.name === "bug")
);

// 4. Post notification if needed
if (criticalIssues.length > 0) {
  await slack.post_message({
    channel: "#alerts",
    text: `⚠️ Found ${criticalIssues.length} critical issues!`,
  });
}

// 5. Return structured summary
return {
  total_issues: issues.length,
  critical_count: criticalIssues.length,
  bug_count: bugIssues.length,
  critical_titles: criticalIssues.map((i) => i.title),
  oldest_issue: issues.sort(
    (a, b) => new Date(a.created_at) - new Date(b.created_at)
  )[0],
};

Advanced Topics

Timeout Configuration

Control execution timeouts at different levels:

// Sandbox-level timeout (E2B)
const client = new MCPClient(config, {
  codeMode: {
    enabled: true,
    executor: "e2b",
    executorOptions: {
      apiKey: process.env.E2B_API_KEY!,
      timeoutMs: 600000, // 10 minutes for long-running tasks
    },
  },
});

// VM executor with custom timeout
const client = new MCPClient(config, {
  codeMode: {
    enabled: true,
    executor: "vm",
    executorOptions: {
      timeoutMs: 60000, // 1 minute default timeout
    },
  },
});

// Per-execution timeout (overrides default)
const result = await client.executeCode(code, 30000); // 30 seconds for this execution

Error Handling

Code execution errors are captured in the result:

const result = await client.executeCode(`
  try {
    const data = await github.get_pull_request({ number: 12345 });
    return data;
  } catch (error) {
    console.error("Failed to fetch PR:", error.message);
    return { error: error.message };
  }
`);

if (result.error) {
  console.error("Execution failed:", result.error);
  console.log("Logs:", result.logs);
} else {
  console.log("Result:", result.result);
  console.log("Execution time:", result.execution_time);
}

Tool Discovery Patterns

Different ways to discover tools based on your needs:

// Quick overview - just names
const names = await search_tools("", "names");
// [{ name: "list_issues", server: "github" }, ...]

// With descriptions for better understanding
const withDesc = await search_tools("file", "descriptions");
// [{ name: "read_file", server: "filesystem", description: "..." }, ...]

// Full schemas for validation
const full = await search_tools("github", "full");
// [{ name: "...", server: "...", description: "...", input_schema: {...} }]

// Check available namespaces
console.log("Available servers:", __tool_namespaces);
// ["github", "filesystem", "slack"]

Resource Cleanup

Always clean up resources when done:

const client = new MCPClient(config, {
  codeMode: {
    enabled: true,
    executor: "e2b",
    executorOptions: {
      apiKey: process.env.E2B_API_KEY!,
    },
  },
});

try {
  // Use the client
  const agent = new MCPAgent({ llm, client, systemPrompt: PROMPTS.CODE_MODE });
  await agent.run("Process data");
} finally {
  // Clean up E2B sandboxes and close MCP sessions
  await client.close();
}

Security Considerations

VM Executor Security

The VM executor uses Node.js vm module for isolation: Allowed:

Standard built-ins: Array, Object, String, Number, Boolean, Date, Math, JSON, RegExp, Map, Set, Promise
Standard functions: parseInt, parseFloat, isNaN, isFinite, encodeURI, decodeURI, setTimeout, clearTimeout
Console methods: console.log, console.error, console.warn, console.info
MCP tools via namespaced functions

Blocked:

require and import statements
process, fs, child_process, and other Node.js modules
eval and Function constructor
Access to host filesystem (except via MCP tools)

Best Practices:

Use VM executor only in trusted environments
Validate and sanitize any user input used in code
Set reasonable timeout limits
Monitor execution logs for suspicious activity

E2B Executor Security

The E2B executor provides stronger isolation:

Code runs in isolated cloud sandbox
Full Linux environment with network access
Sandboxes are ephemeral and destroyed after use
Tool calls are proxied back to host via secure channel
No direct access to host system

Best Practices:

Use E2B for production and multi-tenant systems
Keep API keys secure and rotate regularly
Monitor E2B usage and costs
Set appropriate timeout limits
Review execution logs for debugging

API Reference

MCPClientOptions

interface MCPClientOptions {
  // Enable code mode with boolean (uses default VM executor)
  // OR provide detailed configuration
  codeMode?: boolean | CodeModeConfig;
}

interface CodeModeConfig {
  // Enable code execution mode
  enabled: boolean;

  // Executor type: "vm" (default), "e2b", custom function, or BaseCodeExecutor instance
  executor?: "vm" | "e2b" | CodeExecutorFunction | BaseCodeExecutor;

  // Executor-specific options (type-safe based on executor choice)
  executorOptions?: VMExecutorOptions | E2BExecutorOptions;
}

// VM executor options
interface VMExecutorOptions {
  timeoutMs?: number; // Default: 30000 (30 seconds)
  memoryLimitMb?: number; // Optional memory limit
}

// E2B executor options
interface E2BExecutorOptions {
  apiKey: string; // Required
  timeoutMs?: number; // Default: 300000 (5 minutes)
}

ExecutionResult

interface ExecutionResult {
  // The return value from the executed code
  result: unknown;

  // Console output (log, error, warn, etc.)
  logs: string[];

  // Error message if execution failed, null otherwise
  error: string | null;

  // Execution duration in seconds
  execution_time: number;
}

MCPClient Methods

class MCPClient {
  // Execute code with MCP tool access
  async executeCode(code: string, timeout?: number): Promise<ExecutionResult>;

  // Search available tools (used internally by code execution)
  async searchTools(
    query?: string,
    detailLevel?: "names" | "descriptions" | "full"
  ): Promise<ToolSearchResult[]>;

  // Clean up resources (E2B sandboxes, MCP sessions)
  async close(): Promise<void>;
}

BaseCodeExecutor (for custom executors)

abstract class BaseCodeExecutor {
  constructor(client: MCPClient);

  // Execute code with access to MCP tools
  abstract execute(code: string, timeout?: number): Promise<ExecutionResult>;

  // Clean up resources
  abstract cleanup(): Promise<void>;

  // Get tool namespace information
  protected getToolNamespaces(): ToolNamespaceInfo[];

  // Ensure MCP servers are connected
  protected async ensureServersConnected(): Promise<void>;

  // Create search function for runtime tool discovery
  public createSearchToolsFunction(): SearchToolsFunction;
}

Complete Examples

VM Executor Example

See examples/client/code_mode_example.ts for a complete working example using the default VM executor with the filesystem MCP server. Key points:

Simple configuration with codeMode: true
Uses PROMPTS.CODE_MODE for agent instructions
Demonstrates tool discovery and usage
Shows cleanup with agent.close()

E2B Executor Example

See examples/client/code_mode_e2b_example.ts for a complete working example using E2B cloud sandboxes. Key points:

E2B API key from environment variable
Error handling for missing API key
E2B-specific configuration
Resource cleanup for cloud sandboxes

Troubleshooting

”Code execution mode is not enabled”

Make sure you initialized the client with codeMode: true:

const client = new MCPClient(config, { codeMode: true });

“@e2b/code-interpreter is not installed”

Install the E2B dependency:

yarn add @e2b/code-interpreter

“Script execution timed out”

Increase the timeout:

// For VM executor - set default timeout
const client = new MCPClient(config, {
  codeMode: {
    enabled: true,
    executor: "vm",
    executorOptions: {
      timeoutMs: 60000, // 1 minute default
    },
  },
});

// Or override per execution
const result = await client.executeCode(code, 30000); // 30 seconds

// For E2B executor
const client = new MCPClient(config, {
  codeMode: {
    enabled: true,
    executor: "e2b",
    executorOptions: {
      apiKey: process.env.E2B_API_KEY!,
      timeoutMs: 600000, // 10 minutes
    },
  },
});

Tools not available in code

Make sure MCP servers are configured and connected:

// Check available namespaces
console.log("Available servers:", __tool_namespaces);

// Search for tools
const tools = await search_tools();
console.log("Available tools:", tools);

E2B sandbox issues

Verify API key is correct and valid
Check E2B account has sufficient credits
Review E2B dashboard for sandbox logs
Ensure network connectivity to E2B API

Getting Started

MCP Server

MCP Client

MCP Agents

Development

​Quick Start

​Basic Setup (VM Executor)

​E2B Setup (Cloud Sandbox)

​VM with Custom Timeout

​Architecture

​Flow

​Executor Types

​VM Executor (Default)

​E2B Executor

​Custom Executor

​How Agent Code Works

​1. Tool Namespaces

​2. Tool Discovery

​3. Standard JavaScript

​Complete Example

​Advanced Topics

​Timeout Configuration

​Error Handling

​Tool Discovery Patterns

​Resource Cleanup

​Security Considerations

​VM Executor Security

​E2B Executor Security

​API Reference

​MCPClientOptions

​ExecutionResult

​MCPClient Methods

​BaseCodeExecutor (for custom executors)

​Complete Examples

​VM Executor Example

​E2B Executor Example

​Troubleshooting

​”Code execution mode is not enabled”

​“@e2b/code-interpreter is not installed”

​“Script execution timed out”

​Tools not available in code

​E2B sandbox issues

​Further Reading

Quick Start

Basic Setup (VM Executor)

E2B Setup (Cloud Sandbox)

VM with Custom Timeout

Architecture

Flow

Executor Types

VM Executor (Default)

E2B Executor

Custom Executor

How Agent Code Works

1. Tool Namespaces

2. Tool Discovery

3. Standard JavaScript

Complete Example

Advanced Topics

Timeout Configuration

Error Handling

Tool Discovery Patterns

Resource Cleanup

Security Considerations

VM Executor Security

E2B Executor Security

API Reference

MCPClientOptions

ExecutionResult

MCPClient Methods

BaseCodeExecutor (for custom executors)

Complete Examples

VM Executor Example

E2B Executor Example

Troubleshooting

”Code execution mode is not enabled”

“@e2b/code-interpreter is not installed”

“Script execution timed out”

Tools not available in code

E2B sandbox issues

Further Reading