e2b-codeinterpreter

e2b-codeinterpreter

A forkd parent built from E2B's official code-interpreter template. This is the image Anthropic's tool-use tutorials, OpenAI's code-interpreter examples, and most "let the agent run Python" applications converge on.

Why this recipe is the headline one

• Audience overlap is exact. People who run E2B already use this image; people who want fork-from-warm on top can swap forkd in without re-curating their sandbox tools.
• forkd's Python SDK is E2B wire-compatible. Code that uses from e2b import Sandbox can from forkd import Sandbox and run against this parent unchanged.
• It's tuned for fan-out. Smaller than agent-infra/sandbox (~600 MB), fewer always-running services → faster snapshot + smaller memory.bin → more CoW pages shared at N=100.

What you get

• Python 3 + Jupyter kernel
• numpy, pandas, matplotlib, scipy, scikit-learn, requests preinstalled
• A /code working directory configured for the kernel
• forkd-init.sh + forkd-agent.py wired in as PID 1

Total rootfs: ~600 MB.

Use it

sudo bash recipes/e2b-codeinterpreter/build.sh
sudo bash scripts/host-tap.sh
sudo forkd snapshot --tag ci \
    --kernel ./vmlinux-6.1.141 \
    --rootfs recipes/e2b-codeinterpreter/parent.ext4 \
    --tap forkd-tap0

# Fan out 50 code-interpreter sandboxes
sudo bash scripts/netns-setup.sh 50
sudo -E forkd fork --tag ci -n 50 --per-child-netns --memory-limit-mib 256

# Run pandas analysis in each — picks up the pre-warmed kernel
sudo forkd eval --child forkd-child-7 -- "pandas.DataFrame({'a':[1,2,3]}).sum().to_dict()"

Python SDK (E2B-compatible)

from forkd import Sandbox   # drop-in replacement for `from e2b import Sandbox`

with Sandbox() as sb:
    r = sb.commands.run("python3 -c 'import pandas; print(pandas.__version__)'")
    print(r.stdout)

When to pick this

• You're building an AI code interpreter and want fan-out for per-conversation isolation.
• You already use E2B and want to self-host with significantly faster per-request spawn.
• You want the lightest "agent ready" parent.