CrowdyPy

The official Python SDK for Crowded Kingdoms: typed clients for the platform's one GraphQL API (identity, studio administration, world data, ck-exec) and a native UDP replication core.

CrowdyPy follows the CrowdyJS API surface (the same domains, the same two-token model, the same error codes) the way CrowdyCPP does, with Python names (mintAppToken is mint_app_token). Its realtime path is CrowdyCPP's native replication client, bound with nanobind and shipped inside the wheel. Python never touches a datagram.

v0.10.0: the input log says what it lost (CrowdyJS 18.9.0, CrowdyCPP 0.62.0). Every session from client.input_log.sessions carries missingRecords, the inputs the replication server accepted that never reached the log, which needs the Game API release after v2.40.2. Its docstrings name the new end reasons (paused among them), the two read refusals worth retrying with the same cursor, and how to stop a read nobody will see: cancel its task, which closes the request and frees its server read share. See MIGRATION.md.

v0.9.1: the input log's session filter sends userId as a string, and its reads document when to retry. See the 0.9.1 bullet below; 0.9.0 added the SDK a build compiled against.

v0.8.0: voice, channel audio, wide voxels, and pause and access refusals (CrowdyJS 18.7.0, CrowdyCPP 0.60.0).

• Voice: crowdypy.media's voice helpers (an optional 10-byte header per codec frame, VoicePacketizer, VoiceJitterBuffer), and client.udp.send_channel_audio() with the channel_audio handler, which reaches every member of a channel wherever they are (party and guild voice).
• World Stores: ChunkStore keeps voxel edits the 16x16x16 one-byte grid cannot hold in an overlay, and applies the echo of your own edit once. session.on("voxel_update") and session.on("generic_spatial") reach a game that keeps its own world.
• Refusals: crowdypy.is_app_paused() reads the runtime gate that tokens and the bootstrap carry, and app_paused_of, access_refusal_of and actor_exists_of read the new refusals.
• New calls: users.player_profile(s), app_access.suspend and unsuspend, and exec.restart_type.
• 0.9.0: every build (exec.build, build_status, wait_for_build and the mod builds) carries sdk_version, the ckx-sdk (CLIENT half: crowdy-client-sdk) version the platform compiled it against, whatever the crate names. It needs ck-api v2.40.2, since the build documents select it.
• 0.8.0 needs the ck-api release after v2.39.0, since the token mutations select runtimeGate. Channel audio and the voxel echo need replication server v0.37.0.
• 0.7.0 brought the input log (client.input_log.sessions() and messages(), and replayLoggingEnabled). 0.6.0 brought distance-limited channel messages (udp.send_ranged_channel_message()) and the terms gate (auth.record_player_consents()).
• 0.5.1 made ChunkStore.hydrate() keep every recorded voxel edit (voxelStates) and ensure_around() keep a chunk it already loaded.
• 0.5.0 brought the ck-exec gateway check (client.exec.connect() sends the connect token only to a gateway on the game API's estate, and reports a refusal as Denied) and open grids (game_apps.open_permissions() and set_open_permissions()).
• 0.4 brought the headless Crowdy Studio (crowdypy.studio), player-host observation (crowdypy.player_host), the roots CrowdyCPP covers beyond CrowdyJS, and the strict parity gate (docs/parity-matrix.md has no gaps); 0.3.0 the World Stores, the Game Kit, ck-exec and subscriptions; 0.2.0 native UDP replication (client.udp); 0.1.0 the GraphQL client.

See MIGRATION.md.

Install

pip install crowdypy

Wheels: CPython 3.12+ (one abi3 wheel per platform) and free-threaded CPython 3.14, on Linux x86_64 / aarch64, macOS arm64 / x86_64 and Windows x64. OpenSSL is linked statically; nothing else native is needed. docs/compatibility.md has the details, including the tier versions below.

Each tier of the platform has its own release line, and a build defaults to its own tier's API origin (crowdypy.CROWDY_DEFAULT_TIER says which):

pip skips pre-releases unless a requirement names one, so a prod range never picks up a dev or test build. The same wheels and the sdist are attached to each GitHub release, and pip installs one from its URL (pip install https://github.com/CrowdedKingdoms/CrowdyPy/releases/download/<tag>/<wheel>).

Quick start

import asyncio

import crowdypy

API = "https://ck.dev.crowdedkingdoms.com"


async def main() -> None:
    # The identity client: sign-in, account and studio calls, minting.
    async with crowdypy.AsyncCrowdyClient(http_url=API) as identity:
        await identity.auth.login("player@example.com", "correct-horse-battery")
        minted = await identity.portal.mint_app_token("42")

    # The game client: one per app, at the app's own datacenter.
    async with crowdypy.AsyncCrowdyClient(
        http_url=minted.game_api_url or API, discovery_url=minted.discovery_url
    ) as game:
        game.set_app_token(minted)
        print(await game.server_status.server_with_least_clients())


asyncio.run(main())

Without an event loop, crowdypy.sync.CrowdyClient has the same methods, without await:

from crowdypy.sync import CrowdyClient

with CrowdyClient(http_url=API) as identity:
    identity.auth.login("player@example.com", "correct-horse-battery")
    print(identity.users.me())

Two tokens, two clients

There is one API origin but two tokens, as in every Crowded Kingdoms SDK:

1. Sign-in (auth.login, a magic link, or social/OIDC) yields an identity session token: account, studio administration and minting. It is not accepted for gameplay.
2. Gameplay needs a short-lived app-scoped token per app (portal.mint_app_token(app_id)), which is also the HMAC key for native UDP. Refresh it with client.refresh_gameplay_token() (concurrent callers share one refresh). It is refused with LEGAL_ACCEPTANCE_REQUIRED until the player's agreement to the legal documents and age attestation are stored: show both checkboxes, then call auth.record_player_consents(accept_legal=True, attest_age_of_majority=True).
3. Build one identity client and one client per game. The game client points at the app's datacenter (game_api_url) with discovery_url set, so it can find the app again if that instance stops answering.

Persist them the same way: FileTokenStore.session_path(directory, api_origin) names the session file (one per origin, shared by every game) and FileTokenStore.app_path(directory, app_id) the gameplay one (one per app):

from crowdypy import FileTokenStore

identity = crowdypy.AsyncCrowdyClient(
    http_url=API, token_store=FileTokenStore(FileTokenStore.session_path(state_dir, API))
)

CrowdyPy sends no Origin header, so direct sign-in is served to it; hosted browser sign-in (portal.signIn in CrowdyJS) is a browser flow and is not part of this SDK.

Errors

Every error is a crowdypy.CrowdyError. A GraphQL refusal is a CrowdyGraphQLError with the server's stable code (error.code == "FORBIDDEN") and extensions; HTTP, network and timeout failures have their own classes. CrowdyAppUnavailableError (the app's datacenter is not serving) and CrowdyUserCodeFaultError (a player module faulted; player_fault_of(error) reads the fault) mirror CrowdyJS. A WRONG_DATACENTER refusal moves the client to the app's datacenter and retries once, by itself.

A paused app (its organization ran out of funds, hit a spend cap, or let a subscription lapse) still mints a token, so check crowdypy.is_app_paused(minted.runtime_gate) before entering the world, and tell the player it is paused. Replication then refuses sends with UDP error 33 (APP_PAUSED). crowdypy.app_paused_of(error), access_refusal_of(error) (ACCESS_REVOKED, ACCESS_SUSPENDED with suspended_until, ACCESS_NOT_GRANTED) and actor_exists_of(error) read those refusals. None of them clears by retrying.

Replication

minted = await identity.portal.mint_app_token("42")
async with crowdypy.AsyncCrowdyClient(
    http_url=minted.game_api_url or API, discovery_url=minted.discovery_url
) as game:
    await game.udp.connect(minted)  # assign a server, open the socket
    game.udp.subscribe({"actor_update": lambda n: print(n.uuid, n.chunk, n.payload)})
    await game.udp.send_actor_update((0, 0, 0), my_uuid, pose_bytes)
    echo = await game.udp.send_actor_update_and_wait((0, 0, 0), my_uuid, pose_bytes)

For a game loop, use the connection underneath: send a frame's entities in one call, and read whole batches.

conn = game.udp.connection
conn.send_actor_updates(chunks, uuids, poses, stride=88)  # numpy (n, 3) int64, n x 32 octets
async for batch in conn.batches():
    actors = batch.types == crowdypy.wire.MessageType.ACTOR_UPDATE_NOTIFICATION
    positions = batch.chunks[actors]  # a view, not a copy

crowdypy.replication.ReplicationConnection is the same for code without an event loop: wait() and then poll() from your own loop.

Voxel positions and types are the app's signed 16-bit values, and a voxel state is at most 1,024 bytes (InvalidArgument otherwise, nothing sent). Replication server v0.37.0 echoes every accepted voxel edit back to its sender as a voxel_update.

Voice

send_audio_packet reaches players near a chunk and carries whatever bytes it is given; a game with a voice format of its own keeps it. The voice helpers in crowdypy.media are an optional format that every Crowded Kingdoms SDK shares, byte for byte. Each packet is a 10-byte header (version 1, the codec, a u16 seq, a u32 timestamp in codec samples, the frame's duration and two talk-spurt flags) followed by one codec frame. VoicePacketizer numbers one sender's frames. VoiceJitterBuffer puts each sender's packets back in order and plays them out 60 ms after the first packet of a talk spurt arrived. It reports a frame that never came as a gap, drops one that arrives after its playout time, and holds at most 64 frames a sender.

send_channel_audio reaches every member of a channel wherever they are (party or guild voice), at most 1,024 bytes a packet. The sender needs the channel's send_voice (members_can_speak=True on creation gives it to the member role) and the player's use_voice_chat. There is no echo, and members receive channel_audio:

from crowdypy.media import VoiceCodec, VoiceJitterBuffer, VoicePacketizer

party = VoicePacketizer(VoiceCodec.OPUS, 20)
# Every 20 ms while the player talks (`opus_frame` from your own encoder):
await game.udp.send_channel_audio(channel_id, my_uuid, party.packetize(opus_frame, last=released))
# ...and party.skip() for every 20 ms of silence that is not sent.

voices = VoiceJitterBuffer()
game.udp.subscribe({"channel_audio": lambda n: voices.push(f"{n.channel_id}:{n.uuid}", n.payload)})
for slot in voices.poll():  # at least once a frame, from your audio clock
    if slot.gap:
        conceal(slot.key, slot.frame_ms)  # the frame never came
    else:
        play(slot.key, slot.codec, slot.frame)

No codec ships with CrowdyPy, and the wheel links none: which codec an app uses (Opus, or G.711 µ-law where it has no Opus), and the library that encodes and decodes it, is the app's choice. Where a voice sits in the world (panning, distance attenuation) is the game's too.

World Stores

from crowdypy.codecs import f32, struct_codec, u16
from crowdypy.stores import create_world_session

pose = struct_codec({"x": f32(), "y": f32(), "z": f32(), "yaw": u16()})
session = create_world_session(game, app_id, actor_codec=pose, host=True, save=True)
session.self.join((0, 0, 0), {"x": 0.0, "y": 64.0, "z": 0.0, "yaw": 0})
session.actors.on_join(lambda actor: print("joined", actor.uuid, actor.value))
asyncio.create_task(session.run())  # tick 60 times a second

snapshot = session.actors.snapshot()  # once a frame: every actor at once
states = pose.decode_many(snapshot.state_offsets, snapshot.state_data)  # numpy, no loop

session.chunks is a helper for 16x16x16 chunks with one byte per voxel. An edit that grid cannot hold (a type outside 0-255, a position outside 0-15) is kept in the chunk's overlay(), and voxel_type_at reads it first. The echo of your own set_voxel is not applied a second time. A game with other addressing reads every edit from session.on("voxel_update", ...), which includes those echoes.

Crowdy Studio

from crowdypy.studio import CrowdyStudioController

studio = CrowdyStudioController.for_client(game, app_id=app_id, grid_id=grid_id)
await studio.initialize()  # opens the first project
studio.update_file("SERVER", "src/lib.rs", source)  # autosaves
result = await studio.test_draft()  # build, deploy and switch on the mod
print(result.status, studio.get_state().build_output)

A CLIENT target runs in a host you pass as broker_factory; the browser Studio runs it in a Web Worker.

Performance

Python does no work per datagram. Receive, HMAC verification and decoding happen on CrowdyCPP's network thread, which never takes the GIL; Python sees one batch object per poll. Sends release the GIL, and a batch releases it once for the whole batch. An asyncio loop is woken through a socket, not by polling. Measured against CrowdyCPP on the same machine (benchmarks/README.md): a 200-entity batch costs 1.06x CrowdyCPP's own send path per entity, Python takes in two million verified notifications a second, a notification reaches an asyncio handler 47 µs after it hits the wire (p50), and a 60 Hz loop stays within 0.2 ms of its schedule while receiving 60,000 a second.

Development

uv sync                       # builds the extension (CMake, a C++20 compiler) and installs dev tools
uv run pytest tests/unit
uv run ruff check . && uv run ruff format --check . && uv run mypy

The live suites in tests/e2e/ run against a deployment when the CROWDY_E2E_* variables are set; docs/e2e-coverage.md lists them, how to configure them and the last live runs.

The gates CI runs, all from a clean checkout:

AGENTS.md covers maintenance: re-pinning CrowdyJS and CrowdyCPP, the tier branches and the release path.

License

MIT. Wheels also contain CrowdyCPP (MIT), yyjson (MIT) and OpenSSL (Apache-2.0).