harbor

official tools

Harbor — the agent-evaluation framework and official Terminal-Bench harness (Python). Run agent evals and build RL environments: `harbor run --agent … --model …`. Ships with Python 3.12, git, and the `harbor` CLI.

linux/amd64linux/arm64

by smolmachines updated 6/5/2026

pull

smolvm pack pull registry.smolmachines.com/library/harbor:latest

how to use

Verified steps — with the smolvm CLI installed:

# Pull (Apple Silicon / arm64; use :amd64 on Intel/AMD)

smolvm pack pull registry.smolmachines.com/library/harbor:arm64 -o harbor.smolmachine

# Run a one-off command — ephemeral, the VM is discarded when it exits

smolvm pack run --sidecar harbor.smolmachine harbor --version

# Or run it as a persistent machine — create, start, exec, stop

smolvm machine create harbor --from harbor.smolmachine

smolvm machine start --name harbor

smolvm machine exec --name harbor -- harbor --version

smolvm machine stop --name harbor

Harbor drives coding agents (Claude Code, Codex, OpenHands) — provide the relevant API keys and an execution backend (Docker / Daytona / Modal) at runtime.