jauntbox Get started

Roadmap

Where the trip stands.

Built in plans, each one reviewed and tested before the next starts. Updated 11 November 2026. Plan files live in docs/superpowers/plans; live test results in SMOKE.md.

  • Done built, merged and smoke-tested as noted
  • Next planned next
  • Later not planned yet
  1. Plan 01Done

    Walking skeleton

    • init, up, down, resize, status, night
    • On-box supervisor, limits and safety net
    • Bare jauntbox attaches a herdr session with one workspace per repo
    • Volume and Tailscale identity survive every rebuild

    Live smoke test on 10 Nov: resize in about 2 minutes, self-delete, night sessions and volume survival passed. Four fixes landed from it.

    Plan 01 on GitHub · Smoke test results

  2. Plan 02Done

    Secrets and GitHub

    • jauntbox setup github: private App, rulesets
    • Single-repo tokens, revoked at teardown
    • Secrets pushed to box memory; locked until unlocked
    • On the box, commands answer to jauntctl

    Live smoke test on 10 Nov: App setup, locked box after reboot, permission drift, token revocation and a clean volume passed. Six fixes landed from it.

    Plan 02 on GitHub · Smoke test results

  3. Plan 03aDone

    Privilege split

    • Root ssh from the laptop for jauntbox's own steps
    • Password sudo for you; passwordless only for a few jauntctl commands
    • Your own rootless Docker and Compose, no docker group
    • Metadata endpoint reachable only by root

    Live smoke test on 11 Nov: root ssh install, the sudo split, rootless docker run and Compose, and the metadata block passed. One fix landed from it.

    Plan 03a on GitHub · Smoke test results

  4. Plan 03Done

    Agents in containers

    • jauntctl agent inside a repo checkout in a herdr pane, interactive or unattended
    • One container per run, one repo's token, no Docker socket
    • Metadata endpoint, tailnet and the box itself blocked
    • Wrap-up: the agent writes CONTINUATION.md and pushes before the box goes down

    Built and merged. First live smoke on 11 Nov: image build, Claude in the pane and the isolation checks passed. Run handling, wrap-up and unattended runs are still to be smoke-tested.

    Plan 03 on GitHub · Smoke test results

  5. Plan 04Next

    Your services, reachable by agents

    • You start your services with docker compose up -d, as usual
    • You list the ports agents may use; agents reach them at host.jauntbox:<port>, everything else on the box stays blocked
    • Open an agent’s dev server in your laptop browser, over the tailnet only

    Planned next.

    No plan file yet.

  6. AfterLater

    Further out

    • Dotfiles in containers
    • A clearer terminal UI: progress steps, colours, icons
    • tmux, Gitea/Forgejo, Headscale and NetBird adapters
    • arm64 and other flavours
    • A prebuilt base image for faster boots
    • Packaged installs

    Teams and fleets are a separate project.

Safety ideas on the list but not planned yet: an outbound allowlist (model API, registries and git host only), pinned agent tools with a deny-first permission config, and checks of the home volume's rc files and git hooks on mount.

Get started Read the docs