crewai-python

CrewAI Agent with E2B

This example uses CrewAI's native E2BPythonTool to let an agent execute Python in an isolated E2B sandbox. The agent generates a deterministic dataset and returns statistics plus a checksum that were computed inside the sandbox.

The example uses one persistent tool instance so state is available across retries. The sandbox is always closed explicitly when the crew finishes or raises an error.

Setup and run

1. Install dependencies

Install uv and run:

uv sync

2. Configure API keys

Copy the environment template:

cp .env.template .env

Add:

• an E2B API key;
• an OpenAI API key.

MODEL can be changed to another OpenAI model supported by CrewAI.

3. Run the crew

uv run python main.py

The final answer is a JSON object with the dataset size, mean, population standard deviation, p95, and SHA-256 digest. CrewAI's verbose output also shows the E2B tool call used to compute it.

The deterministic result is:

{
  "count": 10000,
  "mean": 498.0795,
  "population_stddev": 288.902874,
  "p95": 949,
  "sha256": "e7df90c21d14675e7aa60861f67cf9011f49cc1fd58c97bb5622d262fa069662"
}

How sandbox ownership works

persistent=True makes this tool instance reuse one sandbox across calls. The example wraps the crew in try/finally and calls E2BPythonTool.close() instead of relying on process-exit cleanup. Keep the timeout as short as the workload allows.

If you pass an existing sandbox_id, the caller owns that sandbox and the tool will not close it.

Result and error handling

E2BPythonTool returns structured execution data, including stdout, stderr, rich results, and an error field. The task tells the agent to inspect error, correct failed code, and retry rather than treating a failed execution as a valid answer.

Security notes

The sandbox isolates agent-generated code from the host running CrewAI, but tool inputs and outputs should still be treated as untrusted:

• do not put API keys or other secrets in prompts or sandbox files;
• use persistent mode only when the task needs state across calls;
• explicitly close persistent sandboxes;
• use the smallest practical sandbox and execution timeouts.

See the E2B sandbox lifecycle documentation for more details.