asobi Cloud

asobi Cloud is the managed version of the node described everywhere else in these guides. You write the same Lua, your clients speak the same WebSocket and REST protocol, and the callbacks, error codes and Lua API are identical.

What changes is operations. You do not run a container, you do not mount /app/game, and you do not write a sys.config. Environments are created and fed Lua through the asobi CLI, and the operator console is reached from the dashboard rather than by holding a secret.

This guide is the honest version of that trade: the commands that exist, how a bundle actually reaches an environment, and the list of things a cloud tenant cannot do. If the last one rules you out, self-host - the node is Apache-2.0 and complete.

Availability

Invite-only. A new account is created only when the email is on an operator allowlist, on the signup allowlist, or holds an approved waitlist row. Everything else is refused at the point the account would be created.

The public front door is the waitlist on console.asobi.dev: anyone may request access, an operator approves or declines, and admission is single-use. An approval that has been redeemed does not admit a second registration.

Joining a studio that already exists is a different path. An invite from a team owner redeems a single-use token and does not go through the waitlist at all (ADR 0008). One user belongs to one tenant.

What "the same core" does and does not mean

The code is the same. The artefact is not.

A cloud environment runs ghcr.io/widgrensit/asobi_engine, a release that wraps the asobi library published in this repository. Self-hosting runs ghcr.io/widgrensit/asobi, a release built from the same library plus a Lua directory contract. Same modules, same behaviour, two release builds with different boot code: the engine fetches its Lua over HTTP at boot, the image reads it off a mounted directory.

That difference is the source of nearly everything below.

The CLI

Install

curl -fsSL https://raw.githubusercontent.com/widgrensit/asobi-cli/main/install.sh | sh

irm https://raw.githubusercontent.com/widgrensit/asobi-cli/main/install.ps1 | iex

The scripts verify a checksum and install to ~/.local/bin or %LOCALAPPDATA%\asobi\bin; ASOBI_INSTALL_DIR overrides the location and ASOBI_VERSION pins a release. asobi upgrade self-replaces the binary later, verifying a signed checksum first, and refuses to when a package manager owns the install. There is no winget package and no Scoop bucket yet; the release build is wired for both but publishes neither.

Confirm with asobi version.

Commands

asobi env list and asobi destroy <env_id> are the older tenant-wide commands. asobi destroy takes a UUID rather than a name and is best avoided - see Two commands to be careful with.

How a game is chosen

Every environment belongs to a game and a CLI token is tenant-scoped, so an env command needs a game. It resolves in this order:

1. --game <slug> on the command line.
2. The active game from asobi use <slug>.
3. If the tenant has exactly one game, that one.
4. An interactive picker, if stdin is a terminal.
5. Otherwise it is a hard error.

A tenant with two or more games running commands from CI has no terminal for step 4, so pass --game explicitly there.

TLS is not optional

asobi login --saas-url and asobi config set url both refuse anything that is not https, except http to a loopback host for local development. There is no --insecure flag.

Getting started

asobi init mygame --template arena
cd mygame
asobi dev

asobi dev writes .asobi/dev-compose.yml, runs Docker Compose in the foreground and mounts your Lua directory. It never reads credentials and never talks to the control plane, so you can evaluate asobi with no account at all.

The genre starters are arena, basic, chat, turn-based and world, each scaffolding a runnable lua/match.lua. --template defold, godot, unity or backend fetch a full client-plus-backend demo from a pinned repository tag instead.

When you want it hosted:

asobi login
asobi games
asobi use mygame
asobi create prod
asobi deploy prod lua
asobi envs

The last line is not optional politeness. See below.

How Lua reaches an environment

The engine pulls; the control plane never pushes. In order:

1. Collect. The CLI walks the directory you named and takes every .lua file under it, recursively, keying each by its relative path. Assets, JSON and text files are not included - the bundle is Lua and nothing else.
2. Validate locally. Before any network call, three gates run:
   • an entry point (config.lua, or match.lua for single-mode) must sit at the root of the bundle. This is the "you deployed the parent directory" error, and catching it here is the difference between a clear message and an environment that boots with no game modes;
   • every .lua file must compile;
   • config.lua must return a table of mode_name = "script.lua", every named script must be present, and guest_auth must not be a field of that table. The engine reads guest_auth as a top-level global, so a table field is silently ignored.
3. Zip. Deflate, forward-slash entry names, every entry stamped with a fixed timestamp so the bundle is byte-reproducible. The path normalisation matters on Windows, where a backslash entry name would be extracted as a literal file called lua\match.lua and nothing would find your entry point.
4. Upload. POST /internal/cli/envs/<name>/deploy?game=<slug> as application/zip, authenticated with the access token. A 401 refreshes the access token and retries.
5. Gate. The control plane resolves your tenant from the token, the game from the slug within that tenant, and the environment by name within that game. It then refuses with 409 env_not_deployable unless the environment is running, deploying or degraded. Every other state (not_provisioned, provisioning, stopping, stopped, starting, scaling, destroying, destroyed, retired, failed) has no engine that could take a bundle, so the deploy fails loudly rather than being silently dropped.
6. Store. The zip is hashed with SHA-256 and stored, and the generation counter on the environment is claimed before anything is triggered, so two back-to-back deploys cannot collide.
7. Roll. The provisioner rewrites the environment's Kubernetes Secret with the new bundle URL, hash and generation, stamps the generation into the pod template annotations, and re-applies the Deployment. The annotation is what makes Kubernetes actually roll: a Secret change on its own would not.
8. Fetch and verify. The new engine pod reads the bundle URL from its environment, fetches it over the control plane's internal listener authenticated with its own per-environment engine key, verifies the SHA-256, rejects any archive entry that would escape the target directory, and extracts to /tmp/bundle-<generation>.
9. Register. The engine points game_dir at the extracted directory and runs the same config loader a self-hoster's mounted directory goes through - so guest_auth, registration, the mode manifest and every per-mode global behave identically. With no entry point at the root it logs bundle_no_entry_point, lists where your .lua files actually landed, and pins guest_auth off.

Two consequences worth stating plainly

A deploy is a pod replacement, not a hot reload. The Deployment strategy is Recreate: the old pod terminates before the new one starts. In-flight matches and worlds live in memory and do not survive. Self-hosting can hot-reload because the runtime stats a mounted file between ticks; a bundle extracted into /tmp inside a pod never changes for that pod's life, so there is nothing to stat. Design clients to reconnect, and expect a short gap.

A 200 from asobi deploy does not mean the environment is running your new bundle. The CLI prints its success line off the HTTP response, and that response is sent before the Kubernetes work starts - deliberately, because provisioning can outlast an HTTP client's timeout and orphaning live resources against a request that died is worse. The observable outcome is provisioning_status reaching running:

asobi deploy prod lua && asobi envs

A 503 provisioner_unavailable means the bundle was stored but nothing was rolled. Retry the deploy; the stored bundle is not lost.

Authentication

The login flow

asobi login runs an ECDH-encrypted device-code flow:

1. The CLI generates an ephemeral P-256 keypair and posts the public key with a token name (your hostname by default).
2. The control plane returns a session id, a user code and a verification URL. The CLI prints the code and opens a browser.
3. You approve in the dashboard. Approval needs a live dashboard session, so the person approving is someone the control plane already authenticated.
4. The control plane mints tokens, encrypts the payload under the shared secret and stores it against the session.
5. The CLI polls, decrypts locally and writes ~/.asobi/credentials.json.

Because the payload is encrypted end to end, an observer on the polling channel sees nothing usable.

The tokens

The refresh token is bound to a server-issued device secret, 32 random bytes delivered once inside the encrypted payload. Refreshing needs both, and the comparison is constant-time, so a stolen refresh token on its own is useless. The device_fingerprint in your credentials is a hostname and a dashboard label; it is explicitly not a security control.

A refresh re-checks current state before minting: your tenant must still be the tenant recorded at login and that tenant must still be active. A failed check deletes the row, which is what closes the offboarding window.

Scopes are deploy and manage, granted together at approval. There is no scope selection.

Storage and CI

~/.asobi/credentials.json, mode 0600, in a 0700 directory. On Windows the blob is additionally DPAPI-encrypted for the current user, because the file mode creates no ACL there.

ASOBI_ACCESS_TOKEN overrides the stored access token. That is the CI hook:

export ASOBI_ACCESS_TOKEN="$ASOBI_TOKEN"
asobi deploy prod lua --game mygame

Against self-hosting

A self-hoster has no CLI authentication, because there is nothing to authenticate to. No control plane, no tenant, no device flow, no access token. Their credentials are the database password, an ops_secret if they want the ops plane at all, and whatever OAuth, Steam or IAP secrets the game itself uses. They deploy Lua by changing files in a directory.

The console

The operator console is the same nine screens either way. How you get in is not.

Self-hosted, there is one long-lived credential: ops_secret, set by the operator, presented as a bearer token, compared in constant time. There is no default value, so a node that has not configured one rejects every ops request. Over a bearer header it proves all four capability classes, erasure included. A console session opened with it gets every class except erasure, unless console_erasure is set to true.

On cloud, no shared secret ever reaches you:

1. You open the console link on the environment's dashboard page.
2. The control plane runs the same ownership guard every other environment action uses, then maps your team role onto capability classes: owner and admin get read, player_data and config; member gets read and player_data; anything else gets nothing. No role maps to erasure, so no cloud credential can reach it - see below.
3. It signs a token with ASOBI_OPS_TOKEN_SECRET, read out of that environment's own Kubernetes Secret at mint time. The control plane keeps no copy, so a database compromise does not confer the ability to mint.
4. The token lives 900 seconds, which is the ceiling the engine enforces. Minting a longer one would produce a token that verifies nowhere.
5. The dashboard renders a form that posts the token to your environment's /console/session. A body rather than a URL, so the credential never reaches an access log, a referrer or browser history; a form the operator submits rather than script, so it needs no JavaScript and survives a CSP.
6. Your environment verifies the token and opens a session carrying exactly the token's capability classes and expiring no later than the token does. A fifteen-minute credential must not buy a twelve-hour session.

The claim that identifies you is your control-plane user id, not your email, because it lands in that environment's audit rows.

There is no non-interactive path to the ops plane on cloud. A minted token needs a browser session to mint it, so the CI-calls-the-ops-API pattern is a self-hosting capability.

Player erasure is not reachable on cloud. POST /api/v1/ops/players/:id/erase carries the erasure class, and the two credentials that hold it - the ops_secret bearer header, and a console session on a node with console_erasure set - are both operator config a cloud tenant does not have. GET /api/v1/ops/players/:id/export is player_data, so exporting one player's record does work from the cloud console; erasing them does not. If you have to answer deletion requests yourself, that is a reason to self-host. Ask us in the meantime and we will run it against your environment.

What a cloud tenant cannot configure

A cloud environment's sys.config sets four asobi keys, all of them generated and injected by the provisioner: ops_token_secret, env_id, console and guest_verifier_pepper. There is no route, CLI flag, dashboard field or environment-variable passthrough that reaches any other key.

That means the following are unavailable, and this is the honest list rather than a summary:

Identity providers and receipts. oidc_providers, base_url, steam_api_key, steam_app_id, steam_identity, apple_bundle_id, apple_root_cert_path, apple_root_certs, google_package_name, google_service_account_key. A cloud tenant cannot configure Google or Apple sign-in, Steam authentication, or Apple and Google IAP receipt verification at all. This is the largest gap on the list. If your game needs platform sign-in or store receipt validation today, self-host.

Guest auth beyond the toggle. The per-environment pepper is injected for you, so guest_auth = true in your Lua works. guest_verifier_key_id, guest_unlinked_cap, guest_reap_after and guest_reap_interval_ms are not settable, so unclaimed guests are never reaped and only the default soft cap applies. Pepper rotation is not yours either.

Ops-plane shape. ops_secret, console_erasure, console_session_ttl, console_secure_cookie, console_api_base, console_label, console_production. ops_secret and console_erasure are the two that between them put the erasure class out of reach.

Rate limiting and abuse control. rate_limits and all of its buckets, client_gate, client_gate_timeout, client_gate_on_error, trusted_proxies. You run on the defaults and cannot plug in a CAPTCHA or an attestation gate.

WebSocket. ws_allowed_origins, ws_legacy_game_frames, ws_idle_auth_timeout_ms. No cross-site WebSocket-hijacking origin allowlist is available, so the open default stands.

Game structure at the operator layer. game_modes, game_dir (the engine writes it), registration on the operator side, script_registration, extension_restart.

Gameplay bounds. matchmaker tick interval and maximum wait, vote_templates, world_max, world_max_per_player, leaderboard_client_submit, leaderboard_max_boards, auth_cache_ttl_ms, auth_cache_negative_ttl_ms.

Lua runtime dials. max_heap_words, max_reductions_per_ms, reload_mode, config_watch_interval, dev_errors, terrain_providers. A large-world game on cloud is limited to the default terrain provider allowlist.

Clustering. One replica per environment, Recreate strategy. Clustering is not on the table.

Everything outside the asobi application. Database pool size, background job queues, CORS origins (set control-plane-wide, not per tenant), logger level and resilience gates.

What you can set

Everything reachable from the Lua bundle, which is most of what a game actually tunes: the config.lua mode manifest, the per-mode globals (match_size, max_players, strategy, bots, game_type, state_strategy), the world-mode globals (tick_rate, grid_size, zone_size, view_radius, empty_grace_ms, player_ttl_ms), and the two config-layer globals guest_auth and registration.

registration works on cloud precisely because the operator layer leaves it unset - a script value only lands where no operator value exists. guest_auth = true works because the provisioner pre-injects a per-environment pepper for exactly that purpose: the game half is yours, the secret half is not (ADR 0004).

Configuration marks which layer each key belongs to.

Extensions do not run on cloud

An extension is an OTP application added as a dependency of your release and listed in its application set. The engine release is built with a fixed set of applications at image build time, there is no plugin directory, and the bundle loader only ever registers Lua game modes.

Nor can you point an environment at your own image. The column exists in the schema, but no tenant-reachable path writes it.

So: cloud runs the stock engine. If you need an extension - a custom matchmaking strategy in Erlang, a hot loop, your own RPC methods, your own schemas - you need to build a release, and that means self-hosting.

Environments

asobi create <name> --size xs|s|m|l provisions a Deployment, Service, Ingress and per-environment Secret, plus its own PostgreSQL database and login role.

Pricing is flat per environment per month. The current plans are shown in the dashboard at console.asobi.dev, which is the authoritative source; asobi.dev/cloud describes the service but carries no numbers.

Each environment gets a name of the form https://<game>-<env>.<tenant>.asobi.dev with TLS, from a per-tenant wildcard certificate applied ahead of the pod so issuance overlaps startup.

asobi resize updates the row; the new resources apply on the next start, because a Deployment's resources are set at provision time. Stop and start to take them, or wait for the next deploy.

The secrets an environment needs are generated for you and preserved across redeploys, so they stay stable: the distribution cookie, the database password, the guest-auth pepper, the ops-token signing secret and the engine key.

Beyond the CLI, the dashboard carries deletion protection, a retire-then-restore lifecycle with a purge delay, log viewing, live gauges and time-series charts, and environment-down alerting. Team roles, invites and one-tenant-per-user are cloud concepts with no self-hosted equivalent.

Two commands to be careful with

asobi delete <name> does not check deletion protection. The dashboard's destroy button refuses a protected environment and goes through the retire and purge lifecycle; the CLI path tears down the provisioned resources and deletes the row directly. Protection is a dashboard guard, not a CLI one.

asobi destroy <env_id> is worse and should be treated as deprecated. It revokes the environment's API keys and deletes the database row without calling the provisioner, leaving the Deployment, Service, Ingress, Secret and database running with nothing pointing at them. Nothing cleans that up on its own: the control plane's only orphan sweep covers databases, not Kubernetes objects, and it is off in production. Tell us if you have run it, so we can tear the leftovers down by hand. Use asobi delete <name>.

When to self-host instead

Choose cloud when the operational work is the part you do not want: TLS, a database, backups, an ingress, a console you did not have to secure. A game that authenticates with guests or your own accounts, tunes itself from Lua and wants a console its team can share fits cleanly.

Self-host when any of these is true. None of them is a hypothetical.

• You need platform sign-in or store receipt validation. Google, Apple, Steam and IAP verification all need operator config a cloud tenant cannot reach. This is the most common reason.
• You need an extension. Erlang callbacks, custom matchmaking strategies, your own RPC methods or schemas. Cloud runs the stock engine.
• You need to tune the runtime. Rate limits, heap and reduction caps, matchmaker timing, world caps, terrain providers, WebSocket origin allowlists.
• You need hot reload. Mount a directory, edit a file, the runtime picks it up between ticks and in-flight state survives. Every cloud change is a pod replacement.
• You need more than one node. Cloud is one replica per environment.
• You need server-to-server access to the ops plane. An ops_secret works from CI; a minted token needs a browser.
• You have to erase players yourself. The erasure class needs operator config a cloud tenant cannot reach, so deletion requests go through us. Export works from the cloud console; erasure does not.
• You need a specific database, version, port or pool size, or Postgres somewhere you already run it.
• Data residency or procurement says the data stays with you. You do not choose the region an environment runs in; check asobi.dev/cloud for where they run today, and self-host if that is not an answer you can accept.
• The economics do not work. Per-environment pricing against a box you already own is a straightforward comparison, and sometimes the box wins.

You are not locked in either way. A bundle deployed to cloud is the same Lua a self-hosted node reads off a mounted directory, and Exit covers getting your data out.

Next

• Self-hosting - the production compose and the deployment patterns cloud replaces.
• Configuration - every key, and which layer owns it.
• Operator console - the screens, the ops API, what it cannot do.
• Lua scripting - what goes in the bundle.
• Exit - what happens if we stop.