OpenHands kit for Docker Sandboxes

A Docker Sandboxes kind: sandbox kit that boots straight into OpenHands Agent Canvas the browser UI + agent-server on port 8000.

It's a self-contained agent: it owns its image, entrypoint, and credentials, so there's no base-agent coupling and no credential collision. One spec supports Anthropic, OpenAI, and Gemini via LiteLLM.

Architecture

The sandbox runs on docker/sandbox-templates:shell-docker. Your API key stays proxy-managed — the sbx proxy injects the real key on the wire (x-api-key for Anthropic, Bearer for OpenAI, x-goog-api-key for Gemini), so it never enters the sandbox. An egress allow-list bounds outbound traffic.

Quickstart

# 1. Sign in to Docker Hub
sbx login

# 2. Store a key once (Anthropic default; also: openai, google) — never on the CLI
echo "$ANTHROPIC_API_KEY" | sbx secret set anthropic

# 3. Launch — forward port 8000, then open http://localhost:8000
sbx run -p 8000 --kit docker.io/ajeetraina777/sbx-openhands-kits:latest openhands

Or from a local clone / git:

sbx run -p 8000 --kit ./ openhands
sbx run -p 8000 --kit "git+https://github.com/ajeetraina/sbx-kits-openhands.git" openhands

The model defaults to anthropic/claude-opus-4-8 and is pre-seeded into Canvas. Switch providers in Settings > LLM, or recreate with LLM_MODEL overridden and the matching sbx secret set <anthropic|openai|google>.

What the kit does

1. Installs OpenHands Agent Canvas (@openhands/agent-canvas) and the headless openhands runner (base image already ships Node.js 22 + uv).
2. Boots Agent Canvas as the sandbox entrypoint on port 8000 and pre-seeds the model into its settings (Canvas doesn't read LLM_* env vars).
3. Declares Anthropic / OpenAI / Gemini credentials as proxy-managed — the real key is injected on the wire and never stored in the sandbox.

Port 8000 must be forwarded. Launch with sbx run -p 8000 … so the UI is reachable at http://localhost:8000.

Already launched without -p 8000? Forward it after the fact (the ingress on port 8000 binds to all interfaces, so no restart is needed). Use the sandbox name from sbx ls:

sbx ports openhands-sbx-kits-openhands --publish 8000:8000
open http://localhost:8000

Providers

Model (LLM_MODEL)	Provider	Key
anthropic/claude-opus-4-8 (default)	Anthropic	sbx secret set anthropic
openai/gpt-4o	OpenAI	sbx secret set openai
gemini/gemini-2.5-pro	Gemini	sbx secret set google

Automation

A headless runner is also installed for non-interactive use:

openhands --headless --override-with-envs --exit-without-confirmation -t "…"