Agentbox Setup

/agentbox-setup

Box layout (what you're configuring against)

Your user i vscode and you can use sudo to run commands as root.

/workspace is where the user code lives, a per-box git worktree on a fresh agentbox/<box-name> branch (or a tar-piped copy of the host workspace for non-git projects). Run agentbox checkpoint --set-default (similar to docker commit) to save any changes make to the system and workspace so that new boxes will start from a warm state. Everything is wiped on agentbox destroy.

Some special folders:

• Host main repo's .git/ — If the box bind-mounted RW at its identical absolute host path. In-box commits land on the host's branch refs (visible to git log on the host immediately); the box itself carries no SSH/git creds, so git push goes through the host relay (agentbox-ctl git push). The host's working tree is never written to — only refs/objects under .git/. GitHub PR ops (agentbox-ctl git pr create|view|list|comment|review|merge|close|reopen|checkout) flow the same way through host gh; write ops require host confirmation (deny → exit 10), merge and checkout have additional opt-in guards.
• ~/.claude — and similar home folders for coding agents are seeded from the host's ~/.claude on each create so auth, skills, and plugins persist without leaking the host's home dir.
• agentbox.yaml — read by agentbox-ctl from /workspace. Tasks and services declared here are what the supervisor will run.

Exposed ports and services:

• portless - every port with expose: setting in agentbox.yaml, will be exposed not only as a local port but also as a special domain name https://<name>.localhost (so on https) using portless cli and proxy. This will be also mapped to the host where also portless proxy is running so users can access the same service on the same looking url.
• vnc - the webVNC server exposed on 6080 will be proxies to the host on a random port.
• vscode - the vscode server is proxied to the host on a random port.

Goal

Produce a /workspace/agentbox.yaml that captures this project's services, tasks, and box defaults so the in-box supervisor (agentbox-ctl) can boot the workspace deterministically.

agentbox.yaml is declarative. The supervisor reads it on box start, but you don't have to restart the box: after you write the file, agentbox-ctl reload (run from inside the box) makes the already-running supervisor re-read it and immediately run the declared tasks and autostart the services. See step 8.

1. Discover the project

Look at /workspace:

• Top-level manifests: package.json, pyproject.toml / requirements.txt, Cargo.toml, go.mod, Gemfile, composer.json, mix.exs, etc. — these tell you the runtime.
• docker-compose.yaml / docker-compose.yml — often lists the real services the project expects.
• package.json → scripts: look for dev, start, build, test, migrate, seed.
• Makefile / justfile / Taskfile.yaml — alternative task runners.
• Listening ports: grep for listen(, PORT=, framework defaults (3000 for Next.js / Nuxt, 5173 for Vite, 8000 for Django, 8080 for Spring, etc.).
• Database / cache deps to spin up locally (Postgres, Redis, …) — declare them as services if the project doesn't expect them to be external.

2. Pick services and tasks

• Services = long-running. Web servers, watchers, queue workers, databases. restart: on-failure by default.
• Tasks = one-shot. pnpm install, DB migrations, codegen, fixture loaders, install apt packages. Wire dependent services with needs: so they wait for the task to finish successfully.
• Names: must match [A-Za-z0-9_-]+. Task names and service names share a namespace — no collisions.
• No cycles in needs:.
• Always generate a dependency-install task and make it the root of the needs: graph (every service that needs deps gets needs: [install, …]). Future boxes start from a snapshot of the final filesystem so they won't need this, but updates or moving to a cloud provider might need to rebuild the container from scratch. The filesystem can be then later captured by agentbox-ctl checkpoint --set-default. The task must be idempotent: agentbox-ctl re-runs pending tasks on every box stop/start (the daemon dies with the container and is relaunched), so an unguarded install would reinstall on every start. The clean way is the run_once: true field — the supervisor stores a marker keyed by a hash of the command and skips warm boots automatically (the marker lives at /var/lib/agentbox/tasks/<name>, on the box rootfs, captured by checkpoints, never polluting /workspace). Editing the command re-runs it. Detect the package manager from the lockfile — never hardcode pnpm. See the worked example below.
• Add a comment to the beginning of the file to explain what you did and what issues you encountered, so that future run might use this information in case the project evolves and you need to update the agentbox.yaml file.

Stateful services: data persistence & re-seeding (read this for databases)

Declare a containerized dependency with the image: service form — AgentBox generates the docker start-or-run shell (no hand-written docker run … || docker start …). The container runs in the box's dockerd; a published port is reachable from other in-box services at 127.0.0.1:<host port>:

services:
  postgres:
    image:                            # bare string (image: postgres:17-alpine) or a mapping:
      name: postgres:17-alpine
      ports: ["5432:5432"]
      env:
        POSTGRES_PASSWORD: postgres
        POSTGRES_DB: app
      args: "-c max_connections=200"  # string or ["-c","max_connections=200"]
      container_name: app_db          # optional; default = service name
    ready_when: { port: 5432 }
    restart: always

The container is reused by name across box stop/start. (Changing image/env reuses the existing container as-is; docker rm <container_name> + agentbox-ctl reload to apply.) Install the DB client the migrate/seed tasks need (e.g. postgresql-client) in the install task and reach the DB over TCP — don't docker exec the container (nested exec fails with a setns error in a box).

A checkpoint does NOT capture docker-in-docker data. agentbox checkpoint is a docker commit of the box's writable filesystem (the system + /workspace). The in-box dockerd keeps its storage in a separate per-box volume (/var/lib/docker), which is not part of that image — it's fresh on every new box and wiped on agentbox destroy. So a database or cache you run as a docker container (e.g. docker run … postgres) starts empty on every new box created from a checkpoint (every agentbox claude / agentbox create), even though /workspace and any marker files you wrote were restored. (A DB run as a native process with its data dir on the box filesystem — e.g. postgres -D /var/lib/postgresql/data — is captured by the checkpoint, since it lives in the writable layer.)

Consequence for migrate/seed tasks of a containerized DB: do NOT use run_once: true (the marker form). A command-hash marker is correct for deps (they live in /workspace, which the checkpoint captures), but wrong for DB data living in a docker volume: the marker is restored from the checkpoint while the DB is empty, so a marker-guarded seed wrongly skips and the app boots against an empty database. Instead use the run_once: { check: <cmd> } form — the probe runs first and the seed runs unless the probe exits 0, and no marker is written (the DB is the source of truth). Gate on the actual data:

  seed:
    # Re-seed when the DB is empty. The postgres data lives in the in-box docker
    # volume, which is NOT captured by `agentbox checkpoint` — so a box started
    # from a checkpoint has the workspace warm but an empty DB. The marker form
    # would be restored while the DB is blank and wrongly skip; the `check` probe
    # gates on the data itself. Exit 0 = already seeded, skip. Fast no-op once
    # the data is present.
    command: pnpm db:seed
    needs: [install, migrate]
    run_once:
      check: |
        export PGPASSWORD=postgres
        psql -h 127.0.0.1 -p 5432 -U postgres -d app -tAc \
          "SELECT EXISTS (SELECT 1 FROM users LIMIT 1)" 2>/dev/null | grep -q t

Lifecycle nuance (this is why the data check, not a marker, is right):

• Box stop → start (agentbox stop/start): the supervisor daemon dies with the container and is relaunched, so it re-runs all tasks from pending. The per-box docker volume does survive stop/start, so the DB still has data — the data check makes the seed a fast no-op.
• New box from a checkpoint (agentbox claude/create): tasks run and the DB volume is empty → the check fails → the seed runs. Correct.
• Resume after pause (agentbox pause/unpause): the daemon is frozen and thawed, not restarted, so tasks do not re-run at all — nothing to seed, the running DB is untouched.

(Migrations are usually safe to re-run as-is: migration tools track applied migrations in their own table, which on a fresh box is empty, so they simply re-apply. Only the data seed needs the existence check.) Install the DB client the seed/migrate tasks need (e.g. postgresql-client) in the install task — don't docker exec the DB container for these checks (nested docker exec fails inside a box with a setns error); reach it over TCP with the client tools instead.

3. Wire readiness probes (services only)

ready_when: lets the supervisor decide when a service is "ready" (vs. just "running"). Exactly one of these must be present:

• port: 3000 — TCP connect (default host 127.0.0.1; override with host:).
• log_match: "Listening on" — regex matched against stdout/stderr. First match flips the service to ready.
• http: "http://127.0.0.1:3000/health" — GET probe. Optional expect_status: 200 (default: any 2xx).

Tunables: interval_ms (default 500), initial_delay_ms (default 0), timeout_ms (default 60000), on_timeout: kill | mark_unhealthy (default kill — re-enters the restart policy).

Mark the web service with expose:

The box's primary web app (the dev server / Next.js / API the user opens in a browser) should declare:

    expose:
      port: 3000   # the port this service listens on inside the box
      as: 80        # must be 80 — the container port AgentBox publishes

At most one service may set expose:. AgentBox forwards container :80 to 127.0.0.1:<port> and publishes it on the host with portless proxy to a <boxname>.localhost url, so agentbox list/status show it as the box's main URL on every engine (no OrbStack dependency). Set this on the same service whose ready_when: you just wrote (a DB or worker should not get expose:).

If the box hosts a SERVICE AGENT (openclaw), the units are already there

A service agent — agentbox openclaw today — is a daemon AgentBox itself runs. Its supervisor units (an openclaw-onboard one-shot, an openclaw-render, and the openclaw gateway service with ready_when: { http: /healthz } and expose: { port: 18789, as: 80 }) are synthesized from the agent's own catalog row and folded in at boot. Do not write them into agentbox.yaml — a unit of the same name in the workspace file WINS and would replace the one the agent ships.

Two things you may still need to write:

• A different web app on :80. Only ONE service may expose:. If the project has its own web service that must be the box URL, the agent's expose: is dropped with a warning — decide which one the user wants and keep exactly one.
• The openclaw: overlay block (see §11a below), for gateway config the user wants under version control.

4. Restart + backoff

Per service:

• restart: always | on-failure | never (default on-failure).
• backoff: — initial_ms (default 500), max_ms (default 30000; must be >= initial_ms), factor (default 2).

5. (Optional) defaults: block

Sets per-project defaults for agentbox create/claude/code/shell — same shape as ~/.agentbox/config.yaml. CLI flags still override. Common keys:

• box.hostSnapshot (bool) — APFS-clone the host workspace into a per-box scratch dir before seeding /workspace (stabilizes the tar-pipe source).
• box.defaultCheckpoint (string) — checkpoint new boxes start from (normally you set this via agentbox-ctl checkpoint --set-default at the end of setup — see section 9, not by hand).
• box.withPlaywright (bool) — install @playwright/cli globally inside the box.
• box.vnc (bool) — run Xvnc + noVNC on container port 6080.
• box.isolateClaudeConfig (bool) — per-box ~/.claude volume instead of the shared one.
• code.ide — vscode | cursor | auto.
• code.autoTerminals (bool) — auto-generate .vscode/tasks.json with per-service tails.
• browser.default — agent-browser | playwright | both.

Full key list (run on the host): agentbox config list --keys.