Interpreter Tools

Run AI-generated code on your own machine—locally, securely, at lightning speed.

Vercel AI tool provided out of the box.

Interpreter Tools is a drop-in "code-interpreter" backend for AI agents: it spins up lightweight Docker containers, executes untrusted snippets in < 100 ms (with pooling), streams the output, and can be extended to any language by registering a new config object.

Supports pooling, per-session containers, dependency caching, and real-time stdout/stderr—perfect for chat-based tools like GPT function calling, Jupyter-style notebooks, or autonomous agents that need to evaluate code on the fly.

Interactive Shell Example

The most powerful way to experience Interpreter Tools is through our interactive shell example. It provides a chat-like interface where you can:

• Execute commands and code in a secure Docker environment
• Get AI-powered assistance with command execution
• Maintain context of previous commands and their outputs
• Automatically fix errors with AI analysis
• View command history and session information

Try it out:

1# Clone the repository
2git clone https://github.com/CatchTheTornado/interpreter-tools.git
3cd interpreter-tools
4
5# Install dependencies
6yarn install
7
8# Run the interactive shell
9yarn ts-node examples/interactive-shell-example.ts

Check out the video explainer

Presenting Some Cool Use Cases

Here are some examples showcasing different capabilities of Interpreter Tools:

• Interactive Shell - Chat-like interface with AI assistance
• AI Tool Example - Using code execution with Vercel AI
• Ruby Language Support - Extending with new language support
• Direct Engine Usage - Using ExecutionEngine without AI integration

Why Interpreter Tools?

⚡ Sub-100 ms average execution (with container pool & dep-cache). Run untrusted code fast without leaving Node!

🔌 Plug-in language architecture – add a new language by registering one object (see LanguageRegistry). No engine edits required.

📦 Zero-install repeat runs – dependencies are installed once per container and skipped thereafter, saving seconds on every call.

🔒 Docker-level isolation – each snippet executes in its own constrained container (CPU, memory, no-new-privileges).

🖥️ Real-time streaming – stdout/stderr stream back instantly; ideal for REPL-like experiences.

---

Getting Started

Installation

Interpreter Tools is published on npm as interpreter-tools.

Install the package and its dependencies in your Node.js project:

1# Using yarn
2yarn add interpreter-tools ai @ai-sdk/openai
3
4# Or using npm
5npm install interpreter-tools ai @ai-sdk/openai

Quick Start

1. Create a new file example.js in your project:

1const { generateText } = require('ai');
2const { openai } = require('@ai-sdk/openai');
3const { createCodeExecutionTool } = require('interpreter-tools');
4
5async function main() {
6  try {
7    // Create a code execution tool instance
8    const { codeExecutionTool } = createCodeExecutionTool();
9
10    // Use generateText with codeExecutionTool to generate and execute code
11    const result = await generateText({
12      model: openai('gpt-4'),
13      maxSteps: 10,
14      messages: [
15        {
16          role: 'user',
17          content: 'Write a JavaScript function that calculates the sum of numbers from 1 to n and print the result for n=10. Make sure to include a test case.'
18        }
19      ],
20      tools: { codeExecutionTool },
21      toolChoice: 'auto'
22    });
23
24    console.log('AI Response:', result.text);
25    console.log('Execution Results:', result.toolResults);
26
27  } catch (error) {
28    console.error('Error:', error);
29  }
30}
31
32main();

2. Set up your OpenAI API key:

1# Using yarn
2yarn add dotenv
3
4# Or using npm
5npm install dotenv

Create a .env file in your project root:

1OPENAI_API_KEY=your_api_key_here

3. Update your code to use the environment variable:

1require('dotenv').config();
2// ... rest of the code remains the same

4. Run the example:

1node example.js

Direct ExecutionEngine Usage

If you prefer to use the ExecutionEngine directly without the AI integration, here's how to do it:

1. Create a new file direct-example.js:

1const { ExecutionEngine, ContainerStrategy } = require('interpreter-tools');
2
3async function main() {
4  const engine = new ExecutionEngine();
5
6  try {
7    // Create a session with per-execution strategy
8    const sessionId = await engine.createSession({
9      strategy: ContainerStrategy.PER_EXECUTION,
10      containerConfig: {
11        image: 'node:18-alpine',
12        environment: {
13          NODE_ENV: 'development'
14        }
15      }
16    });
17
18    // Execute JavaScript code
19    const result = await engine.executeCode(sessionId, {
20      language: 'javascript',
21      code: `
22const numbers = [1, 2, 3, 4, 5];
23const sum = numbers.reduce((a, b) => a + b, 0);
24const average = sum / numbers.length;
25
26console.log('Numbers:', numbers);
27console.log('Sum:', sum);
28console.log('Average:', average);
29      `,
30      streamOutput: {
31        stdout: (data) => console.log('Container output:', data),
32        stderr: (data) => console.error('Container error:', data)
33      }
34    });
35
36    console.log('Execution Result:');
37    console.log('STDOUT:', result.stdout);
38    console.log('STDERR:', result.stderr);
39    console.log('Exit Code:', result.exitCode);
40    console.log('Execution Time:', result.executionTime, 'ms');
41
42  } catch (error) {
43    console.error('Error:', error);
44  } finally {
45    // Clean up resources
46    await engine.cleanup();
47  }
48}
49
50main();

2. Run the example:

1node direct-example.js

This example demonstrates:

• Creating a session with a specific container strategy
• Configuring the container environment
• Executing code directly in the container
• Handling real-time output streaming
• Proper resource cleanup

TypeScript Support

If you're using TypeScript, you can import the packages with type definitions:

1import { generateText } from 'ai';
2import { openai } from '@ai-sdk/openai';
3import { createCodeExecutionTool } from 'interpreter-tools';
4import 'dotenv/config';
5
6// Rest of the code remains the same

Extending with New Languages (Ruby Example)

Interpreter Tools can support any language that has a runnable Docker image—just register a LanguageConfig at runtime.

1import { LanguageRegistry, LanguageConfig } from 'interpreter-tools';
2
3const rubyConfig: LanguageConfig = {
4  language: 'ruby',
5  defaultImage: 'ruby:3.2-alpine',
6  codeFilename: 'code.rb',
7  prepareFiles: (options, dir) => {
8    const fs = require('fs');
9    const path = require('path');
10    fs.writeFileSync(path.join(dir, 'code.rb'), options.code);
11  },
12  buildInlineCommand: () => ['sh', '-c', 'ruby code.rb'],
13  buildRunAppCommand: (entry) => ['sh', '-c', `ruby ${entry}`]
14};
15
16// Make the engine aware of Ruby
17LanguageRegistry.register(rubyConfig);
18
19// From this point you can use `language: 'ruby'` in `executeCode` or the AI tool.

See examples/ruby-example.ts for a full working script that:

• Registers Ruby support on-the-fly
• Asks the AI model to generate a Ruby script
• Executes the script inside a ruby:3.2-alpine container

Local Development

To set up the project for local development:

1# Clone the repository
2git clone https://github.com/CatchTheTornado/interpreter-tools.git
3cd interpreter-tools
4
5# Install dependencies
6yarn install

Examples

The /examples directory contains several example scripts demonstrating different use cases of the interpreter tools:

AI Tool Example

examples/ai-example.ts Demonstrates how to:

• Use the code execution tool with Vercel AI
• Generate and execute Python code using AI
• Handle AI-generated code execution results
• Process Fibonacci sequence calculation

Run it with:

1yarn ts-node examples/ai-example.ts

Basic Usage Example

examples/basic-usage.ts Shows how to:

• Set up a basic execution environment
• Execute JavaScript code in a container
• Handle execution results and errors
• Use the per-execution container strategy

Run it with:

1yarn ts-node examples/basic-usage.ts

Python Example

examples/python-example.ts Demonstrates how to:

• Execute Python code in a container
• Handle Python dependencies
• Process Python script output
• Use Python-specific container configuration

Run it with:

1yarn ts-node examples/python-example.ts

Python Chart Example

examples/python-chart-example.js Demonstrates how to:

• Generate files from within a Python script (e.g., PNG charts)
• Persist and retrieve these generated artifacts from the container
• Handle file management workflows when running Python code via the engine

Run it with:

1node examples/python-chart-example.js

Python JSON Sort Example

examples/python-json-sort-example.js Demonstrates how to:

• Inject files (in this case JSON) into the container before execution
• Generate additional files inside the container
• Sort and process JSON data using a Python script
• Retrieve the resulting files back to the host

Run it with:

1node examples/python-json-sort-example.js

Shell JSON Processing Example

examples/shell-json-example.ts Demonstrates how to:

• Generate and execute a shell script using AI
• Create directories and JSON files
• Process JSON files using jq
• Handle Alpine Linux package dependencies

Run it with:

1yarn ts-node examples/shell-json-example.ts

Node.js Project Example

examples/nodejs-project-example.ts Shows how to:

• Generate a complete Node.js project structure using AI
• Create an Express server
• Handle project dependencies
• Execute the generated project in a container

Run it with:

1yarn ts-node examples/nodejs-project-example.ts

Shell Example

examples/shell-example.ts A simple example that:

• Creates a shell script
• Executes it in an Alpine Linux container
• Demonstrates basic container configuration
• Shows real-time output streaming

Run it with:

1yarn ts-node examples/shell-example.ts

Ruby Example

examples/ruby-example.ts Shows how to:

• Dynamically register Ruby language support
• Use the AI tool to generate and execute Ruby code

Run it with:

1yarn ts-node examples/ruby-example.ts

Benchmark Examples

examples/benchmark-pool.ts – JavaScript/TypeScript pool benchmark (20 rounds)

1yarn ts-node examples/benchmark-pool.ts

examples/benchmark-pool-python.ts – Python pool benchmark

1yarn ts-node examples/benchmark-pool-python.ts

Average times on a MacBook M2 Pro: JS 40 ms / round, Python 60 ms / round after first run (deps cached).

Usage

The main components of this project are:

1. ExecutionEngine: Manages code execution in containers
2. SessionManager: Manages session state, container metadata, and container history across executions
3. ContainerManager: Handles Docker container lifecycle
4. CodeExecutionTool: Provides a high-level interface for executing code

Basic Usage

1import { createCodeExecutionTool } from 'interpreter-tools';
2
3const { codeExecutionTool }  = createCodeExecutionTool();
4
5// Isolated workspace (default)
6const result = await codeExecutionTool.execute({
7  language: 'javascript',
8  code: 'console.log("Hello, World!");',
9  streamOutput: {
10    stdout: (data) => console.log(data),
11    stderr: (data) => console.error(data)
12  }
13});
14
15// Shared workspace between executions
16const result2 = await codeExecutionTool.execute({
17  language: 'javascript',
18  code: 'console.log("Hello again!");',
19  workspaceSharing: 'shared'  // Share workspace between executions
20});

The workspaceSharing option controls how workspaces are managed:

• isolated (default): Each execution gets its own workspace
• shared: All executions in a session share the same workspace, allowing file persistence between runs

This is particularly useful when you need to:

• Share files between multiple executions
• Maintain state across executions
• Accumulate generated files
• Build up a workspace over multiple steps

Documentation

Architecture Overview

Interpreter Tools is composed of four core layers:

Layer	Responsibility
ExecutionEngine	High-level façade that orchestrates container lifecycle, dependency caching, and code execution.
SessionManager	Manages session state, container metadata, and container history across executions.
ContainerManager	Low-level wrapper around Dockerode that creates, starts, pools, and cleans containers.
LanguageRegistry	Pluggable store of LanguageConfig objects that describe how to build/run code for each language.

All user-facing helpers (e.g. the codeExecutionTool for AI agents) are thin wrappers that forward to ExecutionEngine.

1graph TD;
2  subgraph Runtime
3    A[ExecutionEngine]
4    B[SessionManager]
5    C[ContainerManager]
6    D[Docker]
7  end
8  E[LanguageRegistry]
9  A --> B --> C --> D
10  A --> E

---

Public API

ExecutionEngine

1new ExecutionEngine()
2
3createSession(config: SessionConfig): Promise<string>
4executeCode(sessionId: string, options: ExecutionOptions): Promise<ExecutionResult>
5cleanupSession(sessionId: string): Promise<void>
6cleanup(): Promise<void>
7getSessionInfo(sessionId: string): Promise<SessionInfo>

• SessionConfig – chooses a ContainerStrategy (PER_EXECUTION, POOL, PER_SESSION) and passes a containerConfig (image, env, mounts, limits).
• ExecutionOptions – language, code snippet, optional dependencies, stream handlers, etc.
• ExecutionResult – { stdout, stderr, dependencyStdout, dependencyStderr, exitCode, executionTime }.
  dependencyStdout and dependencyStderr capture any output produced while installing the declared dependencies (e.g. npm, pip, apk) before your code starts executing. They are empty when no dependency phase was required.
• SessionInfo – comprehensive session information including:

  1interface SessionInfo {
  2  sessionId: string;
  3  config: SessionConfig;
  4  currentContainer: {
  5    container: Docker.Container | undefined;
  6    meta: ContainerMeta | undefined;
  7  };
  8  containerHistory: ContainerMeta[];
  9  createdAt: Date;
  10  lastExecutedAt: Date | null;
  11  isActive: boolean;
  12}

Container Metadata

Each container's state is tracked through the ContainerMeta interface:

1interface ContainerMeta {
2  sessionId: string;
3  depsInstalled: boolean;
4  depsChecksum: string | null;
5  baselineFiles: Set<string>;
6  workspaceDir: string;
7  generatedFiles: Set<string>;
8  sessionGeneratedFiles: Set<string>;
9  isRunning: boolean;
10  createdAt: Date;
11  lastExecutedAt: Date | null;
12  containerId: string;
13  imageName: string;          // docker image tag used
14  containerName: string;      // friendly name assigned by ExecutionEngine
15}

This metadata provides:

• Dependency installation status and checksums
• File tracking (baseline, generated, session-generated)
• Container state (running/stopped)
• Timestamps (creation, last execution)
• Container identification
• Image name of the container
• Human-friendly container name

createCodeExecutionTool()

Factory that exposes the engine as an OpenAI function-calling-friendly tool. It validates parameters with Zod and returns the same ExecutionResult.

1import { createCodeExecutionTool } from 'interpreter-tools';
2
3const { codeExecutionTool } = createCodeExecutionTool();
4const { stdout } = await codeExecutionTool.execute({
5  language: 'python',
6  code: 'print("Hello")'
7});

LanguageRegistry

1LanguageRegistry.get(name): LanguageConfig | undefined
2LanguageRegistry.register(config: LanguageConfig): void
3LanguageRegistry.names(): string[]

LanguageConfig fields:

1interface LanguageConfig {
2  language: string;                    // identifier
3  defaultImage: string;                // docker image
4  codeFilename: string;                // filename inside /workspace for inline code
5  prepareFiles(options, dir): void;    // write code + metadata into temp dir
6  buildInlineCommand(depsInstalled): string[];
7  buildRunAppCommand(entry, depsInstalled): string[];
8  installDependencies?(container, options): Promise<void>; // optional pre-exec hook
9}

---

Adding a New Language

1. Create a LanguageConfig object (see the Ruby example).
2. Call LanguageRegistry.register(config) once at startup.
3. Provide a suitable Docker image that has the runtime installed.

No changes to ExecutionEngine are required.

---

Container Strategies

Strategy	Description	When to use
PER_EXECUTION	New container per snippet; removed immediately.	Maximum isolation; slowest.
POOL	Containers are pooled per session and reused—workspace is wiped between runs.	Best latency / resource trade-off for chat bots.
PER_SESSION	One dedicated container for the whole session; not pooled.	Long-running interactive notebooks.

Note

ContainerStrategy.POOL is optimised for fast startup: it re-uses pre-warmed containers and therefore always clears /workspace between executions.
The workspaceSharing: "shared" option is not available in this mode.
If you need persistent state you can either:

1. switch to PER_SESSION (keeps the internal workspace), or
2. mount a host directory via containerConfig.mounts to share specific data/files across pooled runs.

---

Mounts & Environment Variables

containerConfig accepts:

1{
2  image?: string;            // overrides language default
3  mounts?: ContainerMount[]; // { type: 'directory' | 'tmpfs', source, target }
4  environment?: Record<string, string>;
5  cpuLimit?: string;         // e.g. '0.5'
6  memoryLimit?: string;      // e.g. '512m'
7}

Per-Execution Resource Limits

ExecutionOptions let you override CPU and memory for a single run:

1await engine.executeCode(id, {
2  language: 'python',
3  code: 'print(\"hi\")',
4  cpuLimit: '0.5',       // half a CPU core
5  memoryLimit: '256m'    // 256 MB RAM
6});

Under the hood the engine calls container.update({ CpuPeriod, CpuQuota, Memory }) just before execution, so the limits apply even when pooling.

---

Streaming Output

Pass streamOutput: { stdout?, stderr?, dependencyStdout?, dependencyStderr? } in ExecutionOptions to receive data chunks in real time while the process runs.

1await engine.executeCode(id, {
2  language: 'shell',
3  code: 'for i in 1 2 3; do echo $i; sleep 1; done',
4  streamOutput: {
5    stdout: (d) => process.stdout.write(d)
6  }
7});

dependencyStdout / dependencyStderr fire during the dependency-installation phase (e.g. when pip install or npm install runs) before the user code starts. This lets you surface progress logs or errors related to package installation separately from your program's own output.

---

Injecting Files Into the Workspace

Sometimes your code needs additional assets (datasets, JSON files, images, etc.). There are two primary ways to make them available inside the container:

1. Mount a host directory – supply a mounts array in containerConfig when you create a session. This is the easiest way to share many files or large directories.

   1const sessionId = await engine.createSession({
   2  strategy: ContainerStrategy.PER_EXECUTION,
   3  containerConfig: {
   4    image: 'python:3.11-alpine',
   5    mounts: [
   6      {
   7        type: 'directory',                 // always "directory" for now
   8        source: path.resolve(__dirname, 'assets'), // host path
   9        target: '/workspace/assets'        // inside the container
   10      }
   11    ]
   12  }
   13});

2. Programmatically copy / create individual files – use one of the helper methods that work after you have a session:

   1// (a) copy an existing file from the host
   2await engine.copyFileIntoWorkspace(sessionId, './input/data.json', 'data.json');
   3
   4// (b) create a new file from a base64-encoded string
   5await engine.addFileFromBase64(
   6  sessionId,
   7  'notes/hello.txt',
   8  Buffer.from('hi there').toString('base64')
   9);

Both helpers write directly to /workspace, so your script can reference the files with just the relative path.

---

Handling Generated Files

Interpreter Tools automatically tracks new files created in /workspace during an execution:

• ExecutionResult.generatedFiles – list of absolute paths to the files created in that run.
• ExecutionResult.workspaceDir – host path of the temporary workspace directory.

Example:

1const result = await engine.executeCode(sessionId, {
2  language: 'python',
3  code: 'with open("report.txt", "w") as f:\n    f.write("done")',
4});
5
6console.log('New files:', result.generatedFiles);

You can retrieve file contents with:

1const pngBuffer = await engine.readFileBinary(sessionId, 'charts/plot.png');
2fs.writeFileSync('plot.png', pngBuffer);

Keeping generated files after cleanup

By default, calling engine.cleanupSession() or engine.cleanup() removes the containers and their workspaces. Pass true to keep only the files that were detected as generated:

1await engine.cleanupSession(sessionId, /* keepGenerated = */ true);

All non-generated files are removed, the generated ones stay on disk so you can move or process them later. If you mounted a host directory, the files are already on the host and no special flag is needed.

---

Happy hacking!

License

MIT