Build a Platform on Box

This guide is for builders of Lovable-like, Devin-like, or agent-platform products, where your product spawns a Box for each of your own users. What such a platform ultimately needs:

persistent files and context per user or project
an agent that works inside the workspace and can debug anything (the box ships five, see Integrated agents)
a private preview URL of what the user is building
controlled sandbox cost per user

In all cases, you run your own backend and database, and drive Box through the API or SDKs.

The two rules

Always create user boxes with --no-env (noEnv: true). A no-env box receives none of your account’s secrets or credentials and cannot act on your account or other boxes. Pass what the box does need explicitly with --env. See No-env boxes.
Tag each box with per-box environment variables so your backend can identify it.

Instead of remembering --no-env on every call, mark an environment safe for third parties once and point your user boxes at it. The same protection then applies to every box that uses it, and it survives forks and resumes without your backend passing a flag.

box env new users
box env set users --safe-for-third-parties true
box new --environment users --env TENANT_ID=acme

See Environments.

box new --no-env --no-auto-stop --env TENANT_ID=acme

curl -sS -X POST "$BOX_API_BASE/boxes" \
  -H "Authorization: Bearer $BOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"ttlSeconds":null,"noEnv":true,"env":{"TENANT_ID":"acme"}}'

await box.create({
  createBoxRequest: { ttlSeconds: null, noEnv: true, env: { TENANT_ID: "acme" } },
});

box.create(CreateBoxRequest(ttl_seconds=None, no_env=True, env={"TENANT_ID": "acme"}))

Wait for the box to reach ready or idle before running commands in it. This is enforced: a command sent while the box is still provisioning is refused with a retryable 409 box_starting error (or machine_not_running even earlier in provisioning), so wait for ready instead of racing startup. Commands sent too early would also run before your env is applied.

Pattern 1: one always-on box per project

Simplest, perfectly isolated. Spawn a box per user project with --no-auto-stop. Clone or copy the user’s code in, run their dev server, and expose it:

box ssh <id> "cd ~/project && npm run dev -- --host 0.0.0.0 --port 3000 &"
box ssh <id> "host 3000 --private"
# prints https://<subdomain>-3000.on.ascii.dev?_token=...

curl -sS -X POST "$BOX_API_BASE/boxes/$BOX_ID/commands" \
  -H "Authorization: Bearer $BOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"command":"npm run dev -- --host 0.0.0.0 --port 3000 &","cwd":"project"}'

curl -sS -X POST "$BOX_API_BASE/boxes/$BOX_ID/commands" \
  -H "Authorization: Bearer $BOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"command":"host 3000 --private"}'

await box.command({
  boxId,
  commandRequest: { command: "npm run dev -- --host 0.0.0.0 --port 3000 &", cwd: "project" },
});
const hosted = await box.command({ boxId, commandRequest: { command: "host 3000 --private" } });
console.log(hosted.stdout); // https://<subdomain>-3000.on.ascii.dev?_token=...

box.command(box_id, CommandRequest(
    command="npm run dev -- --host 0.0.0.0 --port 3000 &",
    cwd="project",
))
hosted = box.command(box_id, CommandRequest(command="host 3000 --private"))
print(hosted.stdout)  # https://<subdomain>-3000.on.ascii.dev?_token=...

The _token URL is private to whoever you give it to. The preview stays up as long as the box runs. Cost: a default box running 24/7 is about $26/month ($0.00001 per second); a large box is double that. Use this when the project has real traffic or the user pays you enough to cover it.

Pattern 2: one always-on box per user

Like pattern 1, but one box holds all of a user’s projects, each in its own folder on its own port. A small daemon you install in the box multiplexes agent sessions. Works well up to 5 to 10 fullstack projects per user. Cost: about $26/month per user instead of per project. Watch isolation: projects of the same user share a machine.

Pattern 3: stop and resume around usage

Cheapest, isolated, and what most platforms end up with. The box only runs while the user is actively working:

User sends a message. If their box is stopped, box resume it (a few seconds), relaunch your daemon and the preview in parallel, and send the message to the agent.
When the agent finishes, wait a short countdown, then box stop. Stop snapshots the filesystem and pauses billing.

Files, installed packages, and enabled systemd services survive stop and resume. Hand-run processes do not; your daemon restarts them, or run it as an always-on service. See Snapshots.

When the user is done for good

Stop is a pause; delete is the end. DELETE /boxes/{boxId} (box delete, or the ⋯ menu in the dashboard) force-stops the box and permanently deletes the snapshots only it uses. There is no undo and no resume afterwards, so wire it to an explicit “delete my project” action, never to an idle timeout: stopping is what you want when the user might come back. Snapshot data a fork, a resume or a named template still reads is kept. See Snapshots.

When a stop is refused

Stopping saves the box’s disk first. If that save is failing, we refuse the stop and leave the box running rather than throw away the user’s work. You are not billed for that time: the meter pauses on its own from the first failed attempt, so a stuck box never shows up on your invoice and there is nothing to reclaim. See When a stop is refused. Your backend should treat a refused stop as retryable rather than fatal; we retry too, and email the box owner. If you would rather stop it now and accept losing everything written since the last successful snapshot, pass force. It is irreversible, so offer it only after a stop has already failed, and check the last snapshot time first so you know what is being given up.

box stop bx_23456789 --force

curl -sS -X POST "https://ascii.dev/api/box/v1/boxes/bx_23456789/stop" \
  -H "Authorization: Bearer $BOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"force":true}'

await box.stop({ boxId: "bx_23456789", stopRequest: { force: true } });

box.stop(box_id="bx_23456789", stop_request=StopRequest(force=True))

Cost: $1 buys about 27 hours of default machine time, half that on large. A typical user costs $1 to $5 per month, a power user $10 to $20. To pass that cost on, read each box’s own meter for the billing period with GET /boxes/{boxId}/usage (box usage <id>). It works on running and stopped boxes and counts exactly what your balance was charged for that box. See Per-box usage. Two refinements:

Zero preview downtime: publish successful builds to a static host or CDN. The box is only for the agent and dev preview; production traffic never depends on a box being up.
Zero agent latency: run a tool-less copy of the agent on your own server that answers immediately and stalls while the box resumes. See optibox for a working implementation.

Use the integrated agents

Before writing an agent loop, look at what the box already runs. box prompt drives five coding agents (Claude Code, Codex, pi, OpenCode, Prime Agent) with memory, parallel conversations, streaming events, and per-prompt model choice built in. For most platforms it replaces the daemon below. How it works, and every seam to customize it, is on Integrated agents. The tips that matter when you build a product on it:

Your user’s key, not yours. Create the box --no-env and pass the user’s key as a per-box variable: -e ANTHROPIC_API_KEY=…, -e OPENAI_API_KEY=…. Every harness reads it; nothing of your account is on the box. Keys are per box, so users whose keys must stay apart get their own box (patterns 1 and 3), while a box you pay for yourself can be shared across users through conversations (pattern 2).
One conversation per user session. POST /prompt with new: true returns a conversationId; store it next to the session and pass it back as conversationId on every later message. Conversations run in parallel on one box, each with its own memory, and survive stop and resume, so pattern 3 works unchanged: resume the box, prompt the same conversation.
Agent and model picker. A selector in your UI maps directly onto the provider and model fields of POST /prompt. Switching mid-conversation keeps the history, so users can change their mind per message.
Stream to your UI. GET /events with a cursor gives structured prompt, response, and tool-call events, each tagged with its conversationId. Filter with conversation to show one user only their own session.
A stop button. POST /interrupt with conversation stops one user’s turn and leaves the others running.
House rules, tools, MCP servers. The agents read ordinary config files from the box home (AGENTS.md, CLAUDE.md, MCP registrations, skills). Put yours in the template once and every fork starts with them; per-user rules go in the prompt or in per-user boxes. See Customize the harness.
Attachments. box prompt --attach puts the user’s files under ~/attachments on the box and sends images inline to vision models. From a backend, write the file with the file endpoint and name its path in the prompt.

Bring your own harness: the daemon pattern

You do not have to use the built-in agents. If your harness lives in your own infrastructure, or you want full control over the agent loop, put a small HTTP daemon in the box and talk to it directly:

Install it: box scp the binary in, or bake it into your template.
Keep it alive: run it as an always-on systemd service so it survives stop, resume, and fork.
Expose it: host <port> --private gives it a stable HTTPS URL gated by a _token; treat that token as the daemon’s credential and store it in your backend.
Drive it from your backend over plain HTTPS: your harness sends work, the daemon runs it with full machine access and streams results back.

You can also skip the daemon entirely: POST /boxes/{boxId}/commands runs any command in the box, file endpoints read and write data, and first-class endpoints cover lifecycle, prompts, events, desktop, and snapshots. In-box tools such as host are used by running their commands through that same endpoint. A daemon earns its place when you need streaming, concurrency, or lower latency than one-shot commands give you.

Provision from a template

Never install your stack on a fresh box per user. Build it once, then fork:

box new                            # install runtimes, deps, your daemon
box stop <template-id>             # its snapshot is now your template
box fork <template-id> --no-env    # per user, a few seconds

curl -sS -X POST "$BOX_API_BASE/boxes/<template-id>/fork" \
  -H "Authorization: Bearer $BOX_API_KEY" \
  -H "Content-Type: application/json" \
  -d '{"noEnv":true,"env":{"TENANT_ID":"acme"}}'

const forked = await box.fork({
  boxId: templateId,
  forkRequest: { noEnv: true, env: { TENANT_ID: "acme" } },
});

forked = box.fork(template_id, ForkRequest(no_env=True, env={"TENANT_ID": "acme"}))

Per-fork env replaces the per-box variables the fork would inherit from the source, in the CLI (box fork <id> -e KEY=V), the API, and the SDKs alike. Forks inherit the whole filesystem and are usable in seconds at roughly constant cost, whatever the template holds. To update the template: resume it, change it, stop it again. Forks always take the latest snapshot. See Template Boxes.

Sizing

Concurrent boxes track daily active users, not signups. A rough heuristic from platforms running on Box: 10,000 signups is about 100 active users on an average day, so 10 to 100 concurrent boxes, with a launch-day peak several times higher. Size against two separate numbers. Your plan’s concurrent-box cap bounds how many boxes exist at once, and your machine start limits bound how fast you can bring them up. Creating, forking and resuming each count as one start, so pattern 3 spends starts every time a user comes back, not only on signup. See Machine starts.

​The two rules

​Pattern 1: one always-on box per project

​Pattern 2: one always-on box per user

​Pattern 3: stop and resume around usage

​When the user is done for good

​When a stop is refused

​Use the integrated agents

​Bring your own harness: the daemon pattern

​Provision from a template

​Sizing

​Related