E2B 发布 e2b-cookbook：E2B SDK 示例代码与指南合集

Example code and guides for building with E2B SDK.

Read more about E2B on the E2B website and the official E2B documentation.

Examples

Hello World guide

• TypeScript
• Python

Open-source apps

• E2B AI Analyst - analyze your data & create interactive charts
• E2B Fragments - prompt different LLMS to generate apps with UI
• E2B Surf - computer use AI agent powered by OpenAI

Multi-turn reinforcement learning

• OpenEnv Terminus: inspect, run and repair a report in E2B — first reward without a model or GPU, with optional native TRL training wiring.

LLM providers

AI frameworks integrations

Reinforcement learning

• Code rewards for TRL GRPO with E2B — first code rewards without a GPU, then train on JSON-processing tasks with sandboxed Python execution.

Remote execution integrations

Model Context Protocol (MCP)

Example use cases

• Upload dataset and analyze it with Llama 3.3 - Python
• Scrape Airbnb and analyze data with Claude Opus 5 and Firecrawl - TypeScript
• Visualize website topics with Claude Sonnet 5 and Firecrawl - Python
• Next.js app with LLM + Code Interpreter and streaming - TypeScript
• How to run a Docker container in E2B - Python/TypeScript
• Tailcat: encrypted links between sandboxes, and between a sandbox and your laptop - Python/TypeScript
• How to run Playwright in E2B - TypeScript
• Test an app running in a sandbox with a Notte cloud browser, from Playwright over CDP and from the Notte CLI inside the sandbox - Python
• Anchor cloud browser: run code in an E2B sandbox and drive Chromium in Anchor Cloud - TypeScript
• Map custom subdomains to your sandboxes - TypeScript
• Delete old paused sandboxes by age and metadata - Python/Shell
• Feedback analyst agent on Flue, publishing an HTML report from a sandbox - TypeScript
• Warm an E2B sandbox and rerun tests after adding a local regression case - Python
• Teach Hermes an incident-triage playbook and reuse it in a fresh session - Python

Running the examples as a test suite

Every example is exercised nightly against live E2B by tests/run-examples.ts: each one is uploaded into a fresh sandbox, installed with its own toolchain (npm, uv, Poetry, or nbconvert for notebooks) and run. An example passes if it exits 0.

npm install
npm test                      # all of them
npm test -- hello-world       # substring filter, one or a few

You need an E2B_API_KEY plus whichever provider key the examples you are running use - see .env.example. hello-world-js and hello-world-python need only the E2B key, so they are the ones to try first.

What counts as a failure. The suite checks the sandbox, not the model. A provider rate limit or an exhausted quota counts as OK, because the sandbox still built, installed and ran; non-deterministic model behaviour - no chart produced, a malformed tool call, generated code raising inside the sandbox - is reported as skipped. Real failures are the things that are actually broken: missing templates, dependency resolution, wrong entrypoints, auth, retired model ids, sandbox timeouts.

Some examples are deliberately not covered - they need a provider key this repo does not hold, a long-lived server the runner cannot assert on, or an upstream fix. Each exclusion is listed with its reason at the top of tests/run-examples.ts, so a gap is never silent.