horizon/docs/server.md

Hosting a Horizon server: design

Horizon's dedicated server should be the easiest and lightest Minecraft 26.3 server to host. It should be one small native program with no Java and no Python, one command to install it, one config file, and a CLI, a web panel and a control API to manage it. At the same view and simulation distance and with the same players, it should beat vanilla, Paper and Fabric + Lithium on TPS/MSPT, RAM, CPU, chunk generation and start time. Gameplay stays vanilla: there are no Paper-style rule changes unless the owner turns one on.

This page covers what the server does today (measured), the design of horizon-server, and what ships for the first working version (2026-10-23). The owner's guide is docs/hosting.md.

1. Today

How it runs

The dedicated server is the game itself, run headless: horizon-client --resources <prepared> --dedicated [--nogui] [--port N] [--world NAME] [--universe DIR] [--initSettings], started in the server folder (crates/horizon-client/src/game/dedicated.rs).

• Startup. It follows vanilla's Main.main and DedicatedServer.initServer. It reads server.properties and writes it straight back with vanilla's defaults filled in (horizon_client::server, byte-exact with vanilla's Properties.store), and refuses to start until eula.txt says eula=true. It then loads the operator, whitelist and ban lists in vanilla's order and format, and opens <universe>/<level-name> with the same Game as single player, minus the local player.
• Network. Players join through the play-state hub that Open to LAN uses (crates/horizon-client/src/lan). Status ping, legacy ping, compression, login gate (bans, whitelist, IP bans), operator permission levels, the player limit with bypassesPlayerLimit, MOTD and server-icon.png all follow server.properties.
• Main loop. It ticks at 20 TPS on one thread, with chunk generation and lighting on the worker pool. Chunks stay loaded around every player (lan_update_streaming), or around the spawn while nobody is online. It supports pause-when-empty-seconds, player-idle-timeout and the autosave.
• Console (stdin). stop, save-all, save-on/save-off, list [uuids], op/deop, whitelist on|off|list|reload|add|remove, ban, ban-ip, pardon, pardon-ip, banlist, kick and say behave as in DedicatedServer. tempban <player> <duration> [reason] is Horizon's own: a ban that ends by itself (hosting). Every other command goes to the game's own dispatcher as the server (permission level 4). The log follows vanilla's Log4j pattern, in logs/latest.log and on stdout.
• Player names. usercache.json in the universe folder works as vanilla's CachedUserNameToIdResolver. Players are added when they join and kept for a month, and the file keeps the 1000 most recently used, newest first. The console finds a player by name in it, in any case. A name it doesn't hold is asked of the profile service (getByName). If the service doesn't know the name, an offline server uses the offline player of the lower-case name; an online server answers "That player does not exist". tools/test_user_cache_parity.py checks this against vanilla, with a mock profile service.

Measured (2026-10-16, Mac mini M4 Pro, main at bd8dff20, release build)

The server ran in a scratch folder with online-mode=false, level-seed=12345 and view and simulation distance 10. The vanilla server was the pinned 26.3 server.jar on Java 25 with -Xmx4G.

	Horizon (horizon-client --dedicated)	Vanilla 26.3
Start, new world (wall clock, to Done)	11.4 s (Done (9.762s))	12.5 s (Done (2.066s), includes the bundler's first unpack)
Start, existing world (wall clock)	5.8 s (resources load in 1.6 s)	10.7 s (Done (0.446s): the JVM and DataFixerUpper take the rest)
RAM, idle with nobody online (RSS)	1.40–1.51 GB	0.93–0.97 GB
CPU, idle with nobody online	1.2 % of one core	0.6 %
stop to exit	0.6 s	0.6 s
What has to be installed	one 46.8 MB binary + 589 MB of prepared resources	Java 25 + the 62 MB jar
Crates linked	384 (wgpu, winit, cpal, gilrs, arboard, naga…)	n/a

On Linux the full game loads libasound.so.2 (cpal) and libudev.so.1 (gilrs) at startup. wgpu, winit and the Wayland/X11 libraries are loaded only when a window opens. The server build (horizon-engine, section 3) is the same program built without its audio and controllers features, so it needs neither library, and a minimal server image runs it as it is.

The two clear losses are RAM and idle CPU. The dedicated server loads everything the client needs to draw (the HUD atlas, item renderer, block-entity models, panorama, fog and music tables, entity models), and it builds the meshing tables at startup. None of these are used without a window. Loading them is startup work, not tick cost, so it is in scope here (see section 9).

Joining it

• Horizon client: works. A headless Horizon client (--server 127.0.0.1:PORT, the harness mode) logged in, received its 49 chunks at render distance 2, chatted, saw /list, and left. The server's log matched vanilla's lines (logged in with entity id, joined the game, [Not Secure] <Tester> …, lost connection: Disconnected, left the game).
• Vanilla 26.3 client: checked by tools/test_dedicated_server_parity.py. VanillaProtocolClient, a headless client built from vanilla's own codecs, pings and joins Horizon's server and vanilla's from the same world copy. It compares every login, configuration and play packet up to the first chunk batch, the chunk bytes, chat, /list, failing commands, the console lines and the files left behind. Result on main: see "Harnesses on main" below.
• The real vanilla 26.3 client (tools/vanilla_client.py, from the owner's own files, dev-only) joins too. On 2026-10-18 it joined the server build of the game (horizon-engine, section 3) on the Mac mini: python3 tools/vanilla_client.py --name Vanilla --join localhost:PORT. It logged in, drew the terrain around spawn and stayed connected for four minutes, with no warning or error on the server. The tool runs on Linux under vscreen and on macOS from the logged-in screen session, even a locked one, with vsync off.

What fails or is missing

Area	Status
online-mode=true	Works (ServerLoginPacketListenerImpl): a 1024-bit RSA key pair at start ("Generating keypair"), the encryption request, the shared secret and challenge, AES/CFB8, and the session service's hasJoined check (with the address when prevent-proxy-connections is on), with vanilla's log lines and messages. test_online_mode_parity.py runs the same logins against vanilla's server and Horizon's with a mock session service.
Secure chat	Works (SignedMessageChain, LastSeenMessagesValidator, MessageSignatureCache): the chat session update and its profile key, checked against the services' playerCertificateKeys; signed chat, say, me and msg, relayed with their signatures and last-seen lists; enforce-secure-profile (with online-mode) refuses unsigned chat; and vanilla's log lines ([Not Secure], the decoder's warnings, the kicks). Offline servers and players without a chat session chat unsigned, as on vanilla. test_secure_chat_parity.py runs the same signed and unsigned players against vanilla's server and Horizon's, with test keys and a mock services server.
Nether and End	Works: the server holds a level per dimension, and each player is in its own (portals, execute in … tp, respawn). test_server_dimension_players_parity.py checks it against vanilla. How it works, its cost and what's left: server-levels.md.
Refused logins	Logged as vanilla does: Disconnecting <name> (/address): <reason>, then <name> (/address) lost connection: <reason>, and the session errors (Username '<name>' tried to join with an invalid session, Couldn't verify username because servers are unavailable).
Gameplay for remote players	Checked against vanilla's server with a vanilla-codec client (#484, tools/test_dedicated_server_parity.py): a right click with an item on a block uses it once, a dead player's commands run, players on a team without friendly fire don't hurt each other, and a player's projectiles are theirs (their ender pearl takes them). A player's arrow, snowball or egg hits other players like vanilla's, refused by the pvp game rule and by teams without friendly fire (tools/lan_e2e.py arrows; also checked on a dedicated server with two protocol clients, #5785). The recipe book, status effects and block sync while digging (#500) work. Left: a mob a guest kills counts in the host's statistics and advancements, not the guest's (mob_died runs as the host).
Defaults that surprise owners	white-list=true is vanilla 26.3's default, so a fresh server refuses everyone until they are added. That matches vanilla, so the first-run flow (section 3) asks about the whitelist up front.
Management	Only stdin. There is no service, no backups, no status beyond list, no restart after a crash, and a server can't be managed while it runs in the background.
Packaging	The server is the full game client. Resources come from horizon-prepare, which needs a JDK and the official launcher's client install, both awkward on a headless VPS.
Tick times	The dedicated loop never recorded its tick times in the game clock (only the windowed loop did), so /tick query on a dedicated server reported 0 ms. Fixed in PR 2b, which the benchmark needs (section 9).
Scale	Measured against vanilla on equal work (section 9, benchmarks/server-equal-work-2026-10-17/): 20 TPS up to 50 bots, 9 to 11 TPS at 100, where vanilla's watchdog stops it. Horizon is ahead with few players and when exploring; with 25 to 50 players near spawn vanilla's tick is shorter.

Harnesses on main

On the tip of the code PRs (2a to 3, rebased on b4ad6538), tools/parity.py dedicated_server online_mode server_files server_player_combat server_player_damage passed all five. That includes the vanilla protocol client joining Horizon's server and vanilla's from the same world. The full --changed origin/main set (309 harnesses, since the PRs touch the client crate) had passed 60 of 60 when it was stopped for disk space; CI runs it on every pull request.

2. Architecture: a supervisor and an engine

 owner ──► horizon-server (CLI, TUI)          browser ──► web panel           AI ──► MCP (stdio)
                 │                                │                               │
                 └────────── control API: HTTP + JSON on 127.0.0.1, token ────────┘
                                                  │
                                   horizon-server run   (the supervisor: no GPU, no game code)
                  server.toml ──► server.properties ──► │ ◄── stdout: the vanilla log
                  backups, updates, crash restart       │ ──► stdin: console commands
                                                        │ ◄─► engine endpoint: status, TPS/MSPT, players
                                                        ▼
                                   the engine: horizon-client --dedicated   (the game, headless)

horizon-server is a new crate and binary (crates/horizon-server). It is the only program an owner runs. It depends only on the light crates (horizon-core, horizon-protocol, horizon-world, horizon-bench) plus serde, toml, toml_edit and zip, which are already in the lockfile. It doesn't link wgpu, winit, audio or controller crates. It runs the engine as a child process.

The supervisor/engine split stays even after the engine becomes headless, for these reasons:

• Crash restart and diagnostics. The supervisor outlives an engine crash. It keeps the crash report, restarts the engine with backoff, and the panel stays up to say what happened.
• Safe updates. An update swaps the engine binary and restarts. If the new engine fails its start check, the supervisor rolls back to the previous one.
• Parity stays cheap. The engine keeps reading vanilla's own files (server.properties, the four JSON lists, eula.txt) exactly as test_dedicated_server_parity.py checks them today. The clean config is a layer on top and never changes what the engine sees.
• No conflicts. The supervisor touches none of the game code that the tick and integrated-server work (perf-server-1, perf-ticks-2) are changing.

The engine today is horizon-client --dedicated. It needs one small, additive change: a status endpoint. With --control-socket <dir>/.horizon/engine.json (the existing control protocol: TCP on 127.0.0.1 plus a token, on every OS), the dedicated loop answers server-status with TPS, MSPT (average, P50/P95/P99 from the tick clock /tick query uses), players (name, UUID, address, ping, position), loaded chunks and uptime. It answers console <line> with the command's output.

A proxy moving a player between two servers uses more requests on this endpoint (lan/arrivals.rs, game/dedicated/transfer.rs). They are a built-in module, off by default, that the owner turns on in server.toml:

[transfer]
enabled = false              # true: take a proxy's player-* requests; needs Velocity forwarding (forwarding.secret)
expect_seconds = 10          # how long a player the proxy announced is waited for, when it names no time
max_expect_seconds = 60      # the longest the proxy may ask for
max_expected = 64            # players waited for at once
max_freeze_seconds = 30      # the longest a player stays frozen while it is handed over
max_hold_seconds = 60        # the longest a player arriving from another server is held before it is dropped

The supervisor passes the section to the engine as --transfers expect_seconds=10,..., beside --control-socket, and not as a server.properties key: that file stays vanilla's. The section has nothing to do with vanilla's accepts-transfers, which is about the transfer packet a server sends a client. While it is off, every player-* request and transfer-status are answered {"error": "transfers are off on this server"}, the engine keeps no table and does nothing for transfers on a tick.

With it on, the engine refuses to start unless Velocity forwarding is on (forwarding.secret, velocity-forwarding.md). The proxy must be the only way in: a player who joined a source server directly would find there the older copy of itself that a hand-over left behind, and items would double.

Who can ask. The requests do not widen who can reach the engine: they use the control endpoint, a Unix socket or TCP on 127.0.0.1 only (horizon-server starts it on TCP with a token). A proxy on another machine needs something local to each engine to relay for it. The token is all or nothing: whoever can move players can also run any console command.

Answers are one JSON line, {"ok": true, ...} or {"error": "..."}, and a refused request changes nothing. A unique id is written as 32 hexadecimal digits with or without dashes in a request, and as lowercase text with dashes in an answer (123e4567-e89b-12d3-a456-426614174000). A hand-over id is 32 lowercase hexadecimal digits.

transfer-status is what a proxy reads before it sends players to a server. It names no player and changes nothing:

{"ok": true, "players": 12, "held": 1, "expected": 2, "frozen": 0, "online": 13, "max_players": 20, "full": false, "stopping": false}

players are the players placed in the world (a player a moved release took out is no longer counted), held the logins held for a hand-over, expected the expectations still alive, and frozen the players this server is handing over. online is what the player limit counts: the placed and the held players and the logins still joining. full is online at max_players or above: an ordinary login, held or not, is then refused as multiplayer.disconnect.server_full, unless the player is an operator with bypassesPlayerLimit. player-expect counts the expected players too: it is refused once online and expected reach max_players. stopping is true once the server has begun to stop, and no player is taken in any more. A proxy asks it often, so it writes nothing to the log.

player-expect <uuid> [seconds] tells the target server that the player is about to log in: within expect_seconds by default, max_expect_seconds at most (a longer time is cut to it), and at most max_expected players are expected at once. The reply is {"ok": true, "renewed": false, "ttl_seconds": 10}; asking again for the same player renews the time and says "renewed": true. It is refused for a missing or malformed id, a time that is not a whole number of seconds above zero, a player already connected, a full table, a full server ({"error": "the server is full"}) and a server that is stopping ({"error": "the server is stopping"}). Each expectation holds a place until it is used up or runs out: once the places the player limit counts and those of the other expected players reach max-players, a new one is refused as a login would be, so two proxies sending players at once do not fill the server past its limit. As at login, an operator with bypassesPlayerLimit is never refused as full. The places are held against other expectations only: an ordinary login (one nobody expected) sees the vanilla limit alone, so it can take a place an expected player was counting on. That player's login is then refused as server_full, and the proxy aborts the hand-over on the source, where the player goes on as it was. full in transfer-status is likewise the vanilla condition alone. The player's next login uses the expectation up as it starts, once the bans, the whitelist and the player limit let it in, so a login still in configuration when the time runs out (a resource pack prompt) keeps it. A login that then fails before it reaches play (its connection drops in configuration, or the client refuses a required pack) gives the expectation back with its own deadline, so the proxy's next attempt within that time is still expected. An expectation no login used is reported in the log when its time runs out.

A login that used up an expectation is a held player. It gets the whole join sequence and the level around it (chunks, block and light changes, entities, the level's broadcasts), and its acknowledgements are handled (teleport, chunk batch, loaded, keep-alive, ping, tick end, client information), but it is not placed: no join message, no player-list entry for the others, not among the entities they are shown, not simulated, and nothing else it sends is acted on. The world is kept loaded, and generated if need be, around a held player as around a placed one, so a server with nobody else on it sends a held player its chunks. The log says <name> is held: it arrives from another server instead of the join line. A held player nobody takes over is dropped after max_hold_seconds (1,200 ticks by default), with a warning, and its departure is logged as <name> was dropped while held: <reason>, with no "left the game" line. player-export refuses a held player.

player-drop <uuid> forgets the player's expectation, and disconnects the held player without a save or a word to anyone. The reply is {"ok": true, "held": true, "expected": false}: whether a held player was disconnected, and whether an expectation was forgotten. It is refused for an id that is neither held nor expected, and leaves a placed player alone.

player-export <uuid> [freeze] reads a connected player out of the source server. It first closes the player's open menu as vanilla does when a menu closes, in the player's own level (the cursor stack, a crafting table's grid and the inventory's own 2x2 grid go back into the inventory, a trade back to its villager; the client is told to close it and is sent its inventory), then answers:

{"ok": true, "record": {
  "version": 1,
  "data_version": 5023,
  "handover": "9f2c4be0a1d34e6b8c0f1e2d3c4b5a69",
  "uuid": "123e4567-e89b-12d3-a456-426614174000",
  "name": "Alex",
  "dimension": "minecraft:overworld",
  "position": [10.5, 70.0, -3.25],
  "rotation": [90.0, 15.0],
  "velocity": [0.0, 0.0, 0.0],
  "game_mode": 0,
  "player": "<base64>",
  "advancements": { },
  "stats": {"stats": { }, "DataVersion": 5023}
}}

player is the base64 of the uncompressed named NBT (root name empty) of the player compound: what the server writes to players/data/<uuid>.dat before it gzips it, taken from the connected player at that moment, with the values mods attached to the player. advancements and stats are the JSON the server writes to players/advancements/<uuid>.json and players/stats/<uuid>.json, or null on a server that keeps none. version is the version of this record (1); data_version is the world data version of the server that made it (5023 for 26.3). The other fields are a few of the compound's values, for a proxy that does not read NBT. While the player is frozen, the reply also gives "freeze_ticks_left" beside the record: the ticks before the hand-over ends by itself, so that a proxy does not commit on the target too late for its release to come in time.

How a server reads a record (the target's player-import):

• A record of another version, or of another data_version than the server's own, is refused.
• The NBT in player wins over the plain fields: they are copies of its values, for a proxy.
• Fields the reader does not know are ignored, so a later version 1 record may carry more.

The hand-over. freeze starts one: the record gets a handover id, which the later requests name, and the player's own inputs (moves, clicks, chat, commands) stop being acted on, while what answers the server (teleport confirmations, chunk batch acknowledgements, keep-alives, pings, tick ends, client information and chat acknowledgements, and the messages a chat packet says the client has seen, in the order they came) still is. A frozen player is not kicked for floating. Every later export of that player, with or without freeze, answers with the same record, byte for byte, and does not restart the freeze: a proxy that lost the reply asks again. The hand-over ends after max_freeze_seconds (counted in ticks, 600 by default) with a warning in the log, and the id is then forgotten. Without freeze the export is a read: handover is null and the player goes on. The request is refused for a player who is not connected, is held, has not finished joining, is leaving, is riding, is asleep or is dead.

While frozen, the player takes no damage (as vanilla's Abilities.invulnerable: the void and /kill still kill) and picks nothing up (items, arrows, experience orbs), so its record stays what it is until the release. A player who dies while frozen all the same ends the hand-over, with a warning in the log.

player-release <uuid> <hand-over> moved [announce] and player-release <uuid> <hand-over> abort end the hand-over, on the source server. Both name the hand-over the export returned. moved takes the player out of the world because another server has it now: the others see its entity and its player-list entry go, nothing is written to its player file or its records (the other server has the player), the log and the leave event record the departure with the reason moved to another server, and the player's connection is closed with a generic disconnect. The "left the game" line is neither sent nor logged unless announce is added. Nothing is deleted: the player's file here stays as it was. abort lifts the freeze, and the player goes on as if nothing had happened. moved replies {"ok": true, "released": "moved", "announced": false} and abort replies {"ok": true, "released": "abort"}. Both are refused for a player who is not in a hand-over, for another hand-over than the player's current one, and for a player who died while frozen. Once the freeze has run out the hand-over is over, and a late moved is refused: the player has played on since, and the record the other server has is older. The proxy then fails the hand-over on the target (player-drop).

player-import <uuid> <json> applies a record to the held player uuid on the target server, whole or in parts, and lets it into the world:

{"record": { ...the record of player-export... }, "parts": ["inventory", "experience"], "announce": true, "place": true}

parts is optional (all of them by default), announce is optional (true by default) and place is optional (true by default; see below). The parts, in the order the reply lists them, and what each sets:

part	sets
inventory	the inventory, armour, off hand and selected slot
ender_chest	the ender chest
experience	the level, progress, total and seed
health	health and absorption, food, saturation, exhaustion and its timer, and the body's air, fire and freezing
effects	the active effects
game_mode	the game mode, flight and flight speed
position	where the player stands, how it looks, whether it is on the ground, its fall and motion, and the vehicle it rode
spawn_point	the respawn point
recipe_book	the recipe book
advancements	the advancements and the statistics

A part that is not asked for keeps what the held player has, which is what this server loaded for it when it logged in: its own file, or a new player's defaults. An imported inventory or ender chest replaces the held player's whole: a slot the record leaves empty is empty afterwards. The keys of the compound that Horizon does not model, the values mods attached to the player, and the last death, the post effects and the credits, are taken only when every part is. Only the compound of the record is read: the fields above it (position, rotation, velocity, dimension) are for whoever reads the record.

A record is refused, with {"error": "..."} and nothing changed, when its version is not 1, its data_version is not this server's, its uuid is not the request's, it has no handover id (an export without freeze), its compound is not valid base64 NBT, a part's name is unknown, or the player is not held or is leaving. It is refused too when its values are ones no player can have: a position, rotation, motion, health, absorption, saturation, experience progress or fall that is not a number, a position farther than 30,000,512 blocks from the origin (20,000,000 up or down), which vanilla's Entity.load would clamp, a health of zero or less, a health over 1,024 or an absorption over 2,048 (the bounds of max_health and max_absorption), a food exhaustion over 40 (FoodData.addExhaustion's) or an effect amplifier outside 0 to 255. A stack's count out of vanilla's 1 to 99 is read as 1, as for any item. A record whose position is in another dimension than the held player's world is refused when position is among the parts: the client is in one dimension and only a login changes it.

The record's hand-over is bound to the held player at the import: a second import of the same record is taken (a retry), one of another hand-over is refused, and player-confirm and player-drop name it.

The commit point. On success the player is sent what the import changed the way a join sends it (game mode, abilities, held slot, a new position to confirm, the effects, with those it had from this server's own file removed, inventory, recipe book and the player list as it now is). Then, unless "place": false, it is placed as player-confirm places it: its player file, advancements and statistics are written to disk first, each to a new file that is synced and then renamed into place (on Linux and macOS the folder is synced too), then it is placed: simulated, shown to the others and taking part from the next tick, when its first tick sends its health, experience and advancements as a new player's does. Everyone gets its player-list entry, since everyone will be shown its entity; with announce true they also get the join line. The log says <name> placed: it has arrived from another server, and the join line when it is announced. The reply is {"ok": true, "handover": "...", "parts": [...], "placed": true, "announced": true, "recentered": false} only once the file is written, so a proxy sends player-release moved to the source after this reply: from then on a crash of the target, or a power loss, leaves the player's three files there as it arrived. When one cannot be written, the reply is an error and the player stays held and imported, to be confirmed again or dropped.

A lost reply after the commit. A player-import or player-confirm that names the hand-over of a player already placed from it is answered {"ok": true, "placed": true, ..., "retried": true}, with nothing done again: a proxy that lost the reply asks again and learns the commit happened.

A refused moved after the commit. The target then has the player placed and on disk, and nothing undoes that: player-drop takes only a held player. It happens when the freeze ran out between the confirm and the release (the proxy was late), or when an operator's /kill killed the frozen player (nothing else hurts it, and it does not move). The proxy keeps its client on the target and disconnects the source's connection; the source then saves its copy of the player as it is, so the two servers' files differ from that moment. To keep clear of the first case, the proxy confirms only while the export's freeze_ticks_left leaves it time to release, and otherwise drops the arrival and aborts.

Where the player arrives. Only the compound of the record is read, so to put the player somewhere other than the record says (a proxy that knows where its client stands, which the frozen source may not have seen yet), the request takes four optional fields, which win over the compound:

{"record": {...}, "position": [300.5, 64.0, 300.5], "rotation": [90.0, 20.0], "on_ground": false, "velocity": [0.0, -0.08, 0.0]}

position is [x, y, z], rotation is [yaw, pitch] in degrees, on_ground is a boolean and velocity is [x, y, z] in blocks a tick. The pitch is taken as vanilla's Entity.setXRot takes it: modulo 360, then cut to ±90; the yaw is kept as it is. A velocity component over 10 blocks a tick is taken as 0, as Entity.load takes a saved Motion. The velocity is the server's: the arrival's teleport carries no motion, as a vanilla login's does not, so the client starts still, and since a player's client moves itself, its first move replaces it. They belong to the position part: a request that gives one and does not import that part is refused. A position that is not finite, or farther than the record's own may be (30,000,512 blocks from the origin, 20,000,000 up or down), is refused, and so is a rotation or a velocity that is not finite numbers (a rotation past what a float holds too). A record with a vehicle still to restore (RootVehicle) puts the player back on it at the vehicle's own place, so a request that names a position for such a record is refused. The border, the blocks and the chunks are not checked, as a vanilla server does not check the place its player file names. The player is teleported to the position (the client confirms a new teleport id), so an anticheat never sees the arrival as a move, however far it is from the export.

The view follows the imported position. recentered says whether the import moved the player to another chunk than the one its view was built around (only possible with the position part, and only once the player has been sent chunks). When it did, the server moves the view at once, before the position the client confirms: it sends SetChunkCacheCenter for the new chunk and forgets the chunks that are out of reach, as the next tick would have. The chunks around the new place are not sent yet: they come one batch at a time with the next ticks, and the client acknowledges each batch. Whoever reads the player's world (a proxy building a view of it) must therefore wait for new chunk batches after a reply with "recentered": true, as it waited for the first ones; with false the chunks it already has are the ones around the player.

Importing in two steps, so that a hand-over can be cancelled. With "place": false the record is applied and the player's client is sent everything the import changed, but the player stays held: not placed, not simulated, not shown, not announced, and nothing is written. The reply is {"ok": true, "handover": "...", "parts": [...], "placed": false, "announced": <what announce asked>}. Then either:

• player-confirm <uuid> <hand-over> writes the player's files and places it as place: true would have, and announces it if the import's announce said so. What the client did since the import was not acted on, so it is sent its position to confirm again and its inventory. Reply {"ok": true, "placed": true, "announced": true}. It is refused, with nothing changed, for a player who is not held, is leaving, whose record was not imported, or whose record is of another hand-over.
• player-drop <uuid> [hand-over] cancels the arrival: the connection is closed and nothing is saved or said to anyone (the player was never placed, so the world never knew it). Reply {"ok": true, "held": true, "expected": false}. With a hand-over id, it must be the imported record's, so a late drop of an older hand-over leaves a newer one alone; before any import there is no id to name, and the drop goes by the unique id alone (a late drop of that kind can still cut a newer held player that was not imported yet, which loses nothing: the source still has the player frozen, and the proxy aborts there). Once dropped, the player can be neither imported nor confirmed.

A proxy that fails after a successful player-import with place: false, and before it has sent its client the first byte of the hand-over, calls player-drop and the target is as it was. A proxy that dies in between leaves a held player that is dropped by itself after max_hold_seconds, with a warning in the log.

A hand-over, in order. A proxy moving a player from server A to server B, without a loading screen. Both run with [transfer] on and behind the proxy's Velocity forwarding:

1. player-expect <uuid> on B: B holds the player's next login.
2. player-export <uuid> freeze on A: the record and its handover id, and A stops acting on the player's inputs. A lost reply is asked again: the same record comes back.
3. The proxy logs the player in to B on its own connection: B sends the held player the world around the place B loaded for it.
4. player-import <uuid> <json> on B with "place": false and, if the proxy knows where its client stands, the request's position, rotation, on_ground and velocity. With "recentered": true the proxy waits for B's next chunk batches.
5. The proxy switches its client over to B's connection. If anything fails before the client got the first byte of it, player-drop <uuid> <hand-over> on B and player-release <uuid> <hand-over> abort on A, and the player goes on on A as before.
6. player-confirm <uuid> <hand-over> on B: B writes the player's files, then places it. Only after its ok, player-release <uuid> <hand-over> moved on A: the player is out of A's world, and A saves nothing of it.

Every step that is not taken in time is undone by itself: an expectation after expect_seconds, a held player after max_hold_seconds, a freeze after max_freeze_seconds, each with a warning in the log. A moved that comes after the freeze ran out is refused, so the proxy must confirm and release within max_freeze_seconds of the export.

A server that stops. From the /stop command (the console's, a player's or the control endpoint's) or a stop signal on, player-expect, player-import and player-confirm are refused {"error": "the server is stopping"}: a player must not arrive on a server that is about to close. The proxy drops the held player there and aborts on the source. A retry after the commit is still answered, and player-export, player-release and player-drop stay open: they are the way out.

In the log. Each player-* request writes one line with what became of it: INFO and what it did (player-import Alex (<uuid>): ok, parts inventory,experience, placed false, announced true, recentered false, hand-over <id>), or WARN and the reason it was refused (player-export Alex (<uuid>): refused: that player is riding). The hand-over's id is in the lines of the export, the import, the confirm, the release and a named drop, so the lines of one hand-over can be found on both servers. Only the ids, the name when the server knows it, and a few words of the outcome are written: never the record or the request's body, which are large and are the player's. A held player has its own lines instead of the join and leave ones: <name> is held: it arrives from another server, <name> placed: it has arrived from another server and <name> was dropped while held: <reason>.

The headless engine comes after perf-ticks-2's work lands: Resources split into shared and per-game, the game built in Game::new, the integrated server on its own thread. On that base:

1. The dedicated path skips everything only a window uses: the HUD atlas, item renderer, block-entity and entity models, panorama, fog, music, and mesh tables (perf-ticks-2 already stops server-side meshing). This is the RAM and start-time item.
2. The game moves into a library crate, with the window parts behind a client feature, so horizon-engine links no wgpu or winit either. Sound and controllers are already optional (the audio and controllers features, on by default): horizon-engine today is the game built without them (section 3).

3. Packaging

Binaries. The release archives hold horizon-client, horizon-ctl, horizon-prepare and horizon-server. On Linux the release also has horizon-server-linux-x86_64.tar.gz: horizon-server, horizon-prepare and horizon-engine, the game built with --no-default-features (no sound, no controllers), which links neither libasound.so.2 nor libudev.so.1. release.yml builds it after the client archive (xtask package-server) and fails if tools/check_imports.py --server finds either library; CI lints the same build. horizon-server runs horizon-engine when it is beside it or on the PATH, else horizon-client. Build it yourself with cargo build --release -p horizon-client --no-default-features and copy target/release/horizon-client as horizon-engine. macOS and Windows have no such libraries to avoid, so they keep the one archive.

Game files: the owner's own copy, as with prepare. Nothing of Mojang's is redistributed and nothing is downloaded on the owner's behalf. horizon-server looks for prepared resources in this order:

1. [engine] resources in server.toml;
2. resources/ in the server folder;
3. .local/horizon-client/26.3 beside the engine or up to four folders above it: a release folder, the installer's game-files (below), or a checkout.

If none is found, init says where it looked. Today the owner runs horizon-prepare --prepare --server-jar PATH from the release folder (it needs a JDK and the client files, as for playing), or copies a folder prepared on their own computer to the server (scp -r). Planned wrappers:

• On this machine: horizon-server prepare --server-jar PATH runs horizon-prepare on the owner's own server jar plus their launcher install.
• Prepared elsewhere: horizon-server import-resources DIR copies a folder prepared on the owner's own computer, for example by the launcher, to a headless machine. It first checks the folder: horizon-resources.json is there and names Minecraft 26.3, the derived files match their checksums, and the font and texture folders aren't empty; otherwise it lists what is wrong and changes nothing. It copies next to the destination and renames the copy in, so a failure partway keeps the old copy. The destination is the first place the server reads: [engine] resources when set, else resources/ in the server folder when it exists or the server wasn't installed by install.sh, else the installer's game-files/horizon-client/26.3 (where the release's .local/horizon-client points). It ends by saying to start or restart the server.

A server-only prepare (from server.jar alone, without the client jar or a JDK) is the follow-up that removes the last Java step. It needs a list of which prepared files the server really reads, and that list comes out of the headless-engine work.

One-command install (packaging/server/):

• Linux and macOS: curl -fsSL https://raw.githubusercontent.com/horizon-rust/horizon/main/packaging/server/install.sh | sh The script picks the release archive for the OS and architecture (on Linux x86_64 the server archive, on macOS arm64 the game's), downloads it from GitHub Releases, and checks it against the release's signed SHA256SUMS when OpenSSL 3 is installed, else against the SHA-256 GitHub publishes for each asset (docs/security.md). It never asks for sudo. Options: --version vX.Y.Z, --dir (default ~/.local/share/horizon), --bin (default ~/.local/bin), and --archive FILE to install an archive already on the machine (a build of your own; no download and no checksum). A Linux release from before the server archive installs the full game, which links ALSA's and udev's libraries (libasound.so.2, libudev.so.1); minimal server images often lack them (Debian and Ubuntu: sudo apt install libasound2 libudev1, libasound2t64 on Ubuntu 24.04 and later). Then the script lists any that are missing, init warns, and run stops at once with what to install, rather than restarting a game that can't load.
• Windows: irm https://raw.githubusercontent.com/horizon-rust/horizon/main/packaging/server/install.ps1 | iex The script does the same into %LOCALAPPDATA%\Horizon (-Version, -Dir), with junctions instead of links (no admin rights), and adds bin to the user's PATH. It finds OpenSSL 3 on the PATH or in Git for Windows to check the signature.
• Layout. Each release goes to <dir>/releases/<tag> and <dir>/current points at the one in use, so an update or a rollback is one link. <bin>/horizon-server and <bin>/horizon-prepare are two-line launchers that run current. The prepared game files live in <dir>/game-files/horizon-client and each release's .local/horizon-client links there, so updates keep them. The script then prints the next steps (horizon-prepare, horizon-server init).

Docker (packaging/server/Dockerfile, built the same way as the relay's image): a multi-stage build that ends on debian:bookworm-slim as user horizon, with the server build as horizon-engine, so the image needs no ALSA or udev. It uses volume /data for the server folder, takes the owner's prepared game files as a read-only mount on /opt/horizon/.local/horizon-client/26.3, exposes port 25565/tcp, and keeps the control API and the panel on 127.0.0.1 inside the container: docker exec <name> horizon-server status|cmd|backup. On the first start, HORIZON_EULA=true and HORIZON_WORLD, HORIZON_SEED, HORIZON_PORT, HORIZON_WHITELIST, HORIZON_PLAYERS and HORIZON_OP answer init's questions; without the EULA answer the container exits and says why. docker stop sends SIGTERM, which saves and stops cleanly (give it --stop-timeout 180). The image is published only with Ludovic's OK (it's hosted infrastructure).

Services.

• horizon-server service install writes and enables a systemd unit (--user, or a system unit with --system) on Linux, a launchd agent (~/Library/LaunchAgents/dev.horizon.server.<name>.plist) on macOS, or a Scheduled Task at logon on Windows (a real Windows service needs a wrapper and comes later).
• Each runs horizon-server run --dir <folder> through the installer's launcher, so an update applies at the next restart. The supervisor restarts the engine itself, and the service manager restarts the supervisor. service uninstall stops and removes it and leaves the server folder alone.
• service print shows the unit or agent it would write, for owners who'd rather install it by hand.

First run (horizon-server init, interactive on a terminal):

1. EULA. It shows the link to https://aka.ms/MinecraftEULA and asks "Do you agree? [y/N]". Only a yes writes eula=true.
2. World. It asks for the world name (world) and an optional seed. An existing world folder, or a vanilla server folder, is detected and kept: its server.properties is imported (section 4).
3. Port. 25565 by default, checked free right away.
4. Players. It asks whether to use a whitelist (yes by default, as vanilla) and who to add first, then offers to make the first player an operator.
5. Game files. It finds or explains the prepared resources (see above).
6. It writes server.toml, shows the panel address and the commands to start it in the background (horizon-server start) or as a service.

Every question has a flag (--eula, --world, --seed, --port, --whitelist/--no-whitelist, --player NAME, --op NAME, --yes) and an environment variable for Docker (HORIZON_EULA, …; flags win).

4. One config file: server.toml

The owner edits only server.toml. The supervisor writes server.properties from it before every engine start, so the engine stays exactly vanilla. The file is short: typed sections for what owners change, and a [properties] table that passes any other vanilla key through unchanged.

# Horizon server settings (docs/server.md). The server writes them into server.properties before
# every start: edit this file, not that one.

[server]
motd = "A Minecraft Server"
port = 25565
bind = ""                    # listen on every address, or only on this one (server-ip)
max_players = 20
online_mode = true           # check players' Minecraft accounts (false: anyone joins under any name)
whitelist = true             # only whitelisted players and operators can join
enforce_whitelist = false    # kick online players taken off the whitelist
view_distance = 10
simulation_distance = 10

[world]
name = "world"               # the world folder
seed = ""                    # empty: a random seed on the first start
game_mode = "survival"       # survival, creative, adventure or spectator
difficulty = "easy"          # peaceful, easy, normal or hard
hardcore = false

# Any other server.properties key, spelled as vanilla spells it, for example:
# spawn-protection = 16
[properties]

[engine]
path = ""                    # the game program; empty: the one installed beside horizon-server
resources = ""               # the prepared game files; empty: resources/ here, else the usual places
restart_on_crash = true

[backups]
every = "6h"                 # automatic backups while the server runs: 30m, 6h, 1d...; "0" turns them off
keep = 12                    # automatic backups kept (backups made by hand are never deleted)

[updates]
keep = 1                     # other releases kept by updates, besides the one in use and the one rollback goes to

This is the file init writes today, and it also writes [panel] (bind = "127.0.0.1:8130" and allow_public = false, section 6), [api], [transfer] (off; a proxy's hand-overs, section 2) and [discord]. The defaults are vanilla 26.3's (a test checks each one). Unknown keys are errors, as are a typed key set again under [properties] and values vanilla would refuse. There is no backups.before_update setting: an update always backs up the world first (section 7).

• Import. horizon-server init in a folder with a server.properties, or horizon-server import server.properties, reads it with vanilla's own parser (ServerProperties). Keys with typed fields go into their sections; any other key that differs from vanilla's default goes into [properties]. Keys left at their defaults are dropped, so the TOML stays short. The original file is kept as server.properties.imported.
• Changes made in game. The engine writes server.properties itself, for example after /whitelist on. After it stops, the supervisor compares the file with what it wrote and copies any changed keys back into server.toml, so an in-game change survives the next start. Edits go through toml_edit, so the owner's comments and order are kept.
• Changes without opening the file. horizon-server config set KEY=VALUE... (the launcher's Servers page uses it) writes through toml_edit, checks the result as server.toml is checked, and writes nothing if a value is wrong. On a running server it sends what vanilla changes live through its console: difficulty, defaultgamemode and whitelist on|off; server.motd and server.max_players go to the game as the management protocol's serversettings/motd/set and serversettings/max_players/set do, so the server list and the next logins use them at once; the backup settings apply at once because the supervisor reads them each time. Everything else waits for the next start, and the command says which. A hot reload of the view distance needs engine support and comes later.
• server.properties, ops.json, whitelist.json, banned-*.json and world/ keep vanilla's names and formats. A vanilla server folder works as is, and so does going back to vanilla.

5. CLI and TUI

horizon-server init                      first run (section 3)
horizon-server run                       supervisor + engine in the foreground; the console on stdin
horizon-server start | stop | restart    in the background (detached `run`); start also resumes a stopped supervisor
horizon-server status [--json]           up?, uptime, TPS, MSPT, players, RAM, CPU, crashes
horizon-server cmd <command…>            one console command, with its output
horizon-server console                   attach: the live log, and send commands (Ctrl-D detaches)
horizon-server logs [-f] [-n N]
horizon-server op|deop <player>
horizon-server whitelist on|off|list|add|remove [player]
horizon-server players [--json]          who is online, operators, whitelist, bans
horizon-server kick|ban <player> [reason]   pardon <player>   ban-ip <address|player> [reason]   pardon-ip <address>
horizon-server backup [NAME]             backups             restore <backup> [--yes]
horizon-server import <server.properties>
horizon-server config [--json]           config set KEY=VALUE... [--json]
horizon-server service install|uninstall|print [--system]
horizon-server panel                     the panel's address and a one-time login link
horizon-server update [--check] [--version vX.Y.Z] [--timeout S]     rollback [--timeout S]
horizon-server update --archive FILE --version NAME     an archive you have (your own build)
horizon-server mcp                       MCP server on stdio (section 8)
horizon-server import-resources <dir>    copy game files prepared elsewhere into place (checked first)

Planned:
horizon-server prepare --server-jar PATH

• Every command takes --dir (the server folder; default: the current folder).
• While the server runs, management commands go through the control API, so the engine applies them as vanilla's console would and its lists stay right. While it's stopped, op, deop, whitelist, ban, pardon, ban-ip and pardon-ip edit the vanilla files directly with the same code the engine uses (horizon_server_files::user_lists).
• The TUI is horizon-server console (built, console.rs): the recent log and then the live one (the control API's event stream), a one-line header (TPS, MSPT, players, RAM) that updates every second, and a command line. It uses plain ANSI with no TUI framework, and falls back to plain lines when stdout isn't a terminal.

6. Web panel

Scope (MVP, PR 4): status (phase, uptime, version, crashes), live TPS, MSPT, RAM and CPU graphs (the last 10 minutes, sampled once a second), players (list with ping, op, kick), the console (live log plus a command line), and backups (list, back up now, restore with confirmation), with restart and stop. The panel lives in the supervisor, so it is up exactly when the server is: starting a stopped server stays a command (or the service manager's job). Whitelist and bans go through the console for now.

Updates: when a newer release is out, the panel says so with an Update button, which runs the same path as horizon-server update (section 7) and then shows how it went.

Later: bans and IP bans, worlds, server.properties import, logs and crash reports, and modpacks (with the modpack work).

Tech: one HTML page with plain JS and CSS (crates/horizon-server/panel/), embedded in the binary with include_str!. No build step, no framework, no CDN. It uses Horizon's design system tokens (horizon-theme's colours, light and dark) and draws charts on a canvas. It polls the control API's routes once a second (status, metrics?since=, logs?since=), served by the same small HTTP code on std::net, with no async runtime or new dependency. Server-sent events can replace the polling later if a tab ever costs too much.

Auth: local or tailnet only, never an open admin port.

• Bind. The panel listens on 127.0.0.1:8130 by default. panel.bind can name a tailnet address (100.64.0.0/10 or Tailscale's fd7a:115c:a1e0::/48) or another loopback address. Any other address, such as 0.0.0.0 or a public IP, is refused unless panel.allow_public = true is set, and even then the panel logs a warning at every start.
• Who may act. The API token in .horizon/api.json (file mode 0600) is what the CLI and MCP use. Browsers get a one-time login link from horizon-server panel: a code good for 10 minutes, swapped for an HttpOnly; SameSite=Strict session cookie. Anyone who can read the server folder can manage the server, and no one else can. Sessions end when the supervisor stops.
• Protection. The Host header must name the panel itself: its address, localhost on loopback, or a tailnet name (a MagicDNS short name or *.ts.net). That stops DNS rebinding. Changes (POST, PATCH, DELETE) need the X-Horizon: 1 header, which stops CSRF. After 10 failed logins in a minute, logins are refused for a while. The page is served with a strict Content-Security-Policy (nothing from elsewhere) and can't be framed.
• Remote access goes over Tailscale/SSH (ssh -L 8130:127.0.0.1:8130 host), with no TLS or passwords to manage. HTTPS comes for free with tailscale serve, and the doc shows how to set it up.

7. Backups, updates and crash restart

• Backups. The supervisor runs save-off, then save-all flush, then zips the world plus the vanilla files and server.toml into backups/<UTC time>-<name>.zip, then runs save-on. Players stay online. Zip works for everyone, Windows included. It runs on the every schedule, on demand and before an update. It keeps the last keep automatic backups and never deletes one made by hand. A backup is written to a temporary name and renamed when complete, so a crash never leaves a half-written backup that looks finished. A stopped server is backed up straight from its files.

• Restore. restore NAME takes a file name, a unique prefix, or the name given to backup (the newest with it). The server stops. The current world is moved to backups/replaced-<time>/ (never deleted), the backup is unpacked into a temporary folder and checked (level.dat is there), then moved in, and the server starts again if it was running. One command, or one click with confirmation in the panel. restore NAME --dimension minecraft:the_nether (API: dimension in the body; MCP: dimension) brings back only that dimension's folder (dimensions/<namespace>/<path>), the same way, and leaves the other dimensions, level.dat and the players as they are; the folder it replaces goes to backups/replaced-<time>/.

• Stopped game. A stop that reaches the game in game, over RCON or through horizon-server cmd leaves the supervisor, control API and panel alive with no engine process; stop on the supervisor's own stdin (run, a hosting panel's console) still ends both, and while the game is stopped that stdin takes start and stop. horizon-server start or POST /api/v1/start starts the game again; horizon-server stop and the management protocol's server/stop (as vanilla's, the process exits) still shut down both the game and the supervisor. File-backed player-list edits are safe during this stopped state.

• Crash restart. If the engine exits without being asked to stop, the supervisor keeps the last 200 log lines and the engine's crash report (crash-reports/), restarts it after 5 s, and doubles the wait up to 5 minutes. It gives up after 5 crashes in 10 minutes, so it doesn't loop on a broken world, and shows the reason in status and the panel.

• Updates (crates/horizon-server/src/update.rs), for a horizon-server installed by install.sh or install.ps1. Releases ship often (release-flow.md), so updates stay opt-in: a running server looks for a newer release a minute after it starts and then every 6 hours, and only says so, in its log, in status and in the panel (HORIZON_NO_UPDATE_CHECK=1 turns the check off). update --check asks GitHub once. update:

  1. downloads the current Nightly's archive for this computer (the only build Horizon ships since 2026-10-28; GitHub's latest release while there is none; on Linux the server build; --version picks another: a vX.Y.Z release, or a nightly-<date> of the last 7 days) and checks it against its SHA256SUMS, whose Ed25519 signature must check with a key built into horizon-server (docs/release-flow.md, "Signed checksums"), signed for the Nightly channel: an unsigned or tampered build is refused. A Nightly installs as nightly-<date>-<commit>, newer than every vX.Y.Z; another build of the same day is an update too;
  2. unpacks it into <dir>/releases/<tag>, beside the others, with its game files linked to <dir>/game-files as the installers do;
  3. stops the game, backs up the world (backups/<time>-before-<tag>.zip), records the release in use in <dir>/previous and points <dir>/current at the new one;
  4. checks that the new game loads on this system (--help), then starts it;
  5. waits until it has logged Done and answered a status request, for 300 s (--timeout).

  If the new game doesn't load, stops, or isn't up in time, the supervisor goes back by itself: current points at the release before, previous is left as it was, the world is put back from the backup (the one the new game left is kept in backups/replaced-<time>/), and the game starts again on the old release. rollback switches to previous by hand, through the same path (the world is backed up first and kept as it is). The supervisor keeps running as it was: the new horizon-server itself takes over the next time the server starts (stop and start, or the service's restart). A stopped server is switched and its new game checked to load; it starts on the new release next time. One install serves every server folder on the machine, so each one runs the new release from its next start. Once the new game is up (or a stopped server is switched), old releases are deleted from <dir>/releases/: the one in use, the one previous names (where rollback goes), the one the running horizon-server itself came from, and the updates.keep newest of the others (1 by default) stay. A release that can't be deleted (a program still running from it on Windows) is left for the next update.

8. Control API and MCP

• Control API. The control API is the panel's HTTP API, documented as stable (/api/v1/...). The supervisor serves it on 127.0.0.1 at a port of the system's choosing, written with the token to .horizon/api.json. It's plain HTTP/1.1 on std::net (crates/horizon-server/src/http.rs), no runtime.
  • built (PR 2b, 2c): GET status (phase, uptime, crashes, RSS, CPU, and the engine's TPS, MSPT average/P50/P95/P99, players, chunks), GET logs?since=, POST command {line} (returns the output), POST start, POST stop, POST restart, GET backups, POST backups {name}, POST restore {name}, POST update {tag, timeout} (without a tag: the newest release, downloaded first; status has an update object that says how it went), GET metrics?since=;
  • built (server-admin): GET players (online with address, ping, position and op level; ops, whitelist, bans, IP bans), POST players/{kick,ban,pardon,ban-ip,pardon-ip,op,deop}, POST whitelist {enabled}, POST whitelist/{add,remove} (list edits also work while the game is stopped; running edits use vanilla's console command and answer changed and its lines), GET metrics/prometheus (Prometheus text);
  • built (server-admin): GET config (the token hidden), PATCH config {changes} (as config set; not engine.*, panel.*, api.*, nor the RCON, query and management keys, so the API can't widen who reaches the server or change what program runs), DELETE backups?name=, GET backups/download?name= (the zip, streamed);
  • built (server-admin): GET events (server-sent events: log lines, the phase, and the engine's typed join, leave, kick, chat and death events, which the supervisor fetches with server-events SINCE on the engine's control endpoint every 250 ms; JSON when the client doesn't ask for a stream).
• Auth. Every call needs Authorization: Bearer <token from .horizon/api.json> (file mode 0600), and the Host header must name the address it listens on (as the panel's). [api] bind and [api] token fix the address and the token for a program that runs elsewhere (a moderation bot); bind takes loopback or a tailnet address unless allow_public = true, which drops the Host check and warns at every start.
• MCP (built, mcp.rs). horizon-server mcp is an MCP server on stdio, for Claude Code or any MCP client run as the same OS user. It reads the token itself, and its tools map one to one onto the API: status, players, run_command, logs, events, kick, ban, pardon, ban_ip, pardon_ip, op, deop, whitelist, list_backups, backup, delete_backup and restore (both need confirm: true), get_config, set_config, start, restart and stop (needs confirm: true). Tools that send someone away or can't be undone carry MCP's destructiveHint. It's the same idea as the game's horizon-ctl (docs/agent-control.md): an AI can run and diagnose a server, and every action it takes shows in the log.
• RCON. enable-rcon, rcon.port and rcon.password turn on vanilla's remote console, for the hosting panels, Discord bridges and scripts written for it. The supervisor serves it (rcon.rs, std only like http.rs), so it answers while the game restarts (a command then says the game isn't running), and it reads its settings as RconThread.create does: no password or a port outside 1 to 65535 leaves it off with vanilla's warning. The packets are RconClient's: login (3), command (2), the output in packets of 4096 characters, -1 for a refused login, Unknown request <hex> for another type, Error executing: <command> (<reason>) for a command that failed. A command goes to the game as an rcon <line> request on the control endpoint (horizon-server cmd sends console <line>, the same console under the Server source), so the output is what the console prints, its messages joined as RconConsoleSource joins them. Differences, each one named in rcon.rs: the packets are read from the stream, the password is compared in constant time, a connection that hasn't logged in within 30 seconds is closed and at most 16 are open. A command runs as RconConsoleSource runs it: at level 4 from a source named Rcon, whose say reads [Rcon] message and whose feedback reaches the operators in game as [Rcon: message] while broadcast-rcon-to-ops is on (the default), and send_command_feedback is on; the reply to the client always carries the feedback, whatever those say.
• GS4 query. enable-query and query.port turn on vanilla's UDP query, which server lists, status pages and monitoring tools read the MOTD, the version and the players from. The supervisor serves it (query.rs, std only), reads its settings as QueryThreadGs4.create does (a port outside 1 to 65535 leaves it off with vanilla's warning) and speaks QueryThreadGs4's protocol: the handshake (type 9, a token below 2^24 as text), the basic stat and the full stat (a request of 15 bytes), each only with the token. The players come from the game's server-status, read at most once a second, and a full stat is kept 5 seconds, as in vanilla. Differences, each one named in query.rs: nothing answers while the game isn't running, and a session id outside ASCII is echoed as lossy UTF-8.
• Vanilla's own management protocol. Vanilla 26.3 has a server management protocol: JSON-RPC over a WebSocket, set up by the management-server-* keys in server.properties, so tools written for vanilla can manage a Horizon server too. The supervisor serves it (management.rs, the WebSocket in websocket.rs, std only like http.rs; TLS, on by default as in vanilla, through rustls in management_tls.rs, from a PKCS12 keystore, or a self-signed one Horizon makes when none is set), so it answers while the game restarts (server/status says started: false then); calls that change the game go through its console. Authentication, errors and replies follow vanilla's net.minecraft.server.jsonrpc (AuthenticationHandler, Connection, methods/*Service), and rpc.discover gives vanilla's OpenRPC document (management_schema.rs). tools/test_management_protocol_parity.py sends vanilla's server and Horizon's the same messages over TLS (one keystore for both) and compares the answers, and the document with the data generator's json-rpc-api-schema.json.
  • built: the WebSocket and its authentication, TLS, rpc.discover, players, players/kick, server/{status,save,stop,system_message}, and the four lists (allowlist/*, bans/*, ip_bans/*, operators/*), whose changes the game applies itself (its control endpoint's management request, in game/dedicated/management.rs), so the source, expiry, log lines and kicks are vanilla's;
  • notifications: the supervisor pushes them to every connected client as vanilla's JsonRpcNotificationService does (a JSON-RPC request without an id, its one parameter in an array), through a queue and a thread so a slow client never holds the game's output. Built: server/started (the game's Done) and server/stopping (when the supervisor or the game's own stop begins the shutdown, once per run), server/saving and server/saved (around every world save of the running game: save-all, the autosave, the pause when empty; a save the game makes while it stops is after the last one the supervisor hears of), and players/joined and players/left (from the game's typed events, within a quarter of a second), and operators/*, allowlist/*, bans/* and ip_bans/* added/removed: each of the game's lists tells a watcher what add, remove and clear do, as vanilla's StoredUserList tells its NotificationService, so a change by the console, a command in the game or the protocol itself is told alike (clear once per entry); the event stream carries them as operators_added, bans_removed... too;
  • game rules: gamerules lists every rule the game has as a typed rule ({key, type, value}), gamerules/update sets one (checked against its type and bounds with vanilla's error texts, and set, logged and told on every update, as vanilla's does) and notification/gamerules/updated follows every change, by that method or by /gamerule (Game::set_game_rule); the parity script compares both servers on all of it;
  • server/activity: a login attempt (the game records login_activity in its login gate, as ServerActivityMonitor.reportLoginActivity is called when a login starts), told at most once in thirty seconds and lost otherwise, as the monitor clears its flag whatever it did; and server/status, the heartbeat: every status-heartbeat-interval seconds (none at 0, the first after as many) the status server/status answers, up or not, to whoever is connected;
  • planned: the rest of serversettings/* and the world-upgrade notifications (after WorldUpgrader's server side).

9. Benchmark plan (bots against vanilla and Paper)

The goal is the published claim: "runs 50–100 players on hardware where vanilla struggles". It must be reproducible by anyone, and show the same gameplay on both sides.

• Servers:

  • Horizon (horizon-server);
  • vanilla 26.3 (the owner's server.jar);
  • Paper (a jar the owner downloads; Paper's defaults are kept, and noted as non-vanilla);
  • optionally Fabric + Lithium (via the pinned tools/benchmark-deps/mods.lock.json).

  Java servers run on Java 25 with -Xms = -Xmx (the same heap for all, 4 GB by default), and the heap is reported.

• Same conditions:

  • the same seed (12345) and a fresh world, plus a second pass on a pre-generated copy of the same world;
  • view-distance 10 and simulation-distance 10;
  • online-mode=false and the same server.properties;
  • one server at a time on the same machine, with the bots on another machine on the LAN, or pinned to other cores when they share one (the report records which).

• Bots: horizon-bots (PR 5), a small Rust load client on horizon-protocol. It handles handshake, offline login and configuration, accepts teleports, answers keep-alives and pings, acknowledges chunk batches as vanilla's client does, and sends movement at 20 Hz. The bots play in creative and fly 24 blocks above their spawn point: they never touch terrain, so they need no copy of the world, and no server can refuse a move for walking into a block. Each behaviour is deterministic per bot from a seed:

  • idle: hovers above spawn;
  • walk: flies to random points within 48 blocks of spawn at sprint speed (5.6 blocks/s);
  • explore: flies straight outwards at creative flying speed (22 blocks/s), which forces chunk generation;
  • mixed: 60 % walk, 20 % explore, 20 % idle.
  • fight: survival bots (gamemode=survival, force-gamemode=true, allow-flight=true) that track the entities the server sends, walk to the nearest living one and attack it every 5 ticks, vanilla's bare-hand cooldown; with --pvp true (the default) they attack each other too, which needs pvp=true on the server. The report counts attacks.

  A build behaviour (breaking and placing blocks) needs the bots to read terrain and comes later. Runs use 1, 10, 25, 50 and 100 bots, joining 1 per second. Bots are needed because 100 real vanilla clients can't run on one machine.

• Measured (once a second, for 5 minutes after the last bot joins, after a 1-minute warm-up):

  • TPS and MSPT (average, P50/P95/P99), from each server's own /tick query (vanilla's ServerTickRateManager; Horizon implements the same command and clock), so every server measures itself the same way;
  • RSS and CPU (process tree, horizon_bench::rss_bytes/cpu_seconds);
  • chunk generation: chunks per second while the explore bots run, and time until each new bot has its full view loaded;
  • start time: cold and warm, to Done;
  • bot-side: chunk arrival, keep-alive round trip, and disconnects (a run with disconnects is reported as failed).

• Output. tools/server_bench.py drives it: a fresh server per run from the same server.properties, Done, the bots, a warm-up, then the measurement. It writes a horizon-server-benchmark JSON report (the machine, commit, Java version, settings and every run) and prints a comparison table. Results are published like the client benchmark, with the machine, the versions and the exact commands. TPS is derived from /tick query's average (20 while it fits in 50 ms), since the command reports times, not the rate.

• First run (PR 5, on a busy shared Mac mini, so the numbers are noisy). With 1 bot Horizon ticks in 3.5 ms against vanilla's 11.2 ms and uses 1.6 GB against 3.6 GB. At 10 and 25 mixed bots Horizon falls to about 10 TPS where vanilla keeps 20, and at 50 both land at about 15 TPS. The work isn't equal yet:

  • Vanilla pulls fast players back when it lags (moved too quickly), which Horizon never does, so vanilla's explore bots generated less terrain.
  • Vanilla's chunk sending yields under load.
  • Horizon sent each nearby walk bot 1.5–4 times as many chunks.

  Horizon sent 2–4 times the chunks per second in every multi-bot run.

• Equal-work run (2026-10-17, benchmarks/server-equal-work-2026-10-17/, on the ThinkPad with the bots on the Mac mini). Horizon now checks moves and paces chunks as vanilla does (#586), the bots wait for the chunk they stand in, an --explore-speed pass keeps both servers at the same rate, and vanilla runs with gamerule locator_bar false (Horizon had no locator bar when this ran; since #692, merged the same day, it sends one to vanilla clients). Results:

  • Memory: Horizon's RSS is 1.9 to 2.6 times smaller in every run (vanilla's heap is fixed at 4 GB).
  • Walking near spawn: Horizon's tick is shorter with 1 and 10 bots (5.5 against 18.3 ms, 11.5 against 16.8), and longer with 25 and 50 (20.7 against 16.0, 31.4 against 20.5). Both keep 20 TPS up to 50.
  • Exploring at the same rate: Horizon's tick is about half of vanilla's (13.5 against 29.2 ms at 10 bots, 20.2 against 37.0 at 25).
  • At 100 bots: vanilla's watchdog stops it in every pass; Horizon runs at 9 to 11 TPS.

  Two server costs were found and fixed on the way: every chunk section was encoded once per player (#623), and every packet woke its connection's writer thread (#649). Together they took the 50-bot walking tick from 281 ms to 31 ms. At 25 walking bots most of what remains is the engine's streaming and block entity ticks, both of which grow with players (the README has the profile).

• Mixed bots, as installed (2026-10-18, benchmarks/server-mixed-2026-10-18/, the first published run for the working version). It used the v2026.10.18.15 release as install.sh installs it, against vanilla 26.3, with 1, 10, 25 and 50 mixed bots and vanilla's default game rules on both (Horizon has the locator bar now). The servers ran on the ThinkPad and the bots on the Mac mini. Results:

  • Both kept 20 TPS in every run.
  • Horizon's RSS was 3.7 to 5.2 times smaller (1.0 GB against 4.0 GB at 50).
  • Horizon's average tick was shorter with 1, 10 and 25 bots (2.4 against 17.9 ms, 12.4 against 34.5, 24.0 against 31.6) and longer with 50 (35.2 against 29.7). Its P99 was lower at every count.
  • Horizon also did more work: it kept up with the exploring bots, so it generated 1.9 to 2.9 times the new terrain a second.

• Mixed bots, equal work (2026-10-18, benchmarks/server-equal-mixed-2026-10-18/). This is the same run with the load made equal. The explore bots fly at 0.8 blocks a second, so every bot follows the same path on both servers and none waits for terrain. A 300-second warm-up lets both finish the bots' first views before the measurement. server_bench.py now prints a work table that compares each count's chunks a second across servers and flags a run that isn't equal. Horizon ran as horizon-engine, the server build, installed from its archive. Results, on the ThinkPad again (the desktop didn't wake):

  • The work was equal at 1 and 10 bots. At 25 the chunks sent a second were equal (158.6 against 157.6), and vanilla still sent 2.5 more first-time chunks a second while it caught up.
  • Tick: 1.8 against 11.3 ms at 1 bot, 4.7 against 34.3 at 10, 11.7 against 50.5 at 25 (vanilla 19.8 TPS). At 50 Horizon ran at 23.8 ms. Vanilla's watchdog stopped its server 30 s after the last bot joined (a 60-second tick). This is one pass, to repeat on a bigger machine.
  • Horizon's RSS was 4.1 to 4.9 times smaller (1.03 GB at 50).
  • Served in full, vanilla's walking bots got 5400 to 6100 chunks each, against 640 to 1100 in the first run. That is why vanilla held 50 bots there and not here.

• Four servers, equal work (2026-10-19, benchmarks/server-four-way-2026-10-19/). The same load against vanilla, Paper (build 205, its defaults except the connection throttle, which refused bots that share an address) and vanilla on Fabric with Lithium. Horizon was release v2026.10.19, which spawns mobs and ticks blocks around every player as vanilla does (#950, #931), so it now does vanilla's work away from spawn too. Results, on the ThinkPad (the desktop didn't wake):

  • The work was equal on all four at every count (chunks a second within 4 %).
  • Horizon's RSS was 3.8 to 4.8 times smaller (0.72 to 1.11 GB). It started in 8.2 to 8.5 s, against 14.7 to 22.4 s for the Java servers. From 10 bots up it used the least CPU.
  • Average tick: Paper was ahead of Horizon at every count (10.3 against 13.7 ms at 1 bot, 28.3 against 30.9 at 25, 46.9 against 47.8 at 50). Fabric + Lithium was ahead at 1 and 10 bots (12.5 and 20.2 against 13.7 and 22.9), behind at 25, and fell to 17.2 TPS at 50. Vanilla was slowest, and its watchdog stopped it at 50.
  • Horizon's P99 was the lowest from 10 bots up (73 ms at 50, against Paper's 110 and Fabric's 174), and its worst single P99 the lowest at every count.

• Ownership. perf-server-1 owns making the tick faster. This work only measures it, files what it finds, and fixes startup, RAM and packaging.

10. Working version for 2026-10-23

Goal: horizon-server installs with one command, starts from a clean config, serves Horizon and vanilla 26.3 clients, backs up and restores, has a CLI and a basic web panel, and has a first published benchmark against vanilla.

In:

1. horizon-server crate and binary, horizon-server-files (the vanilla files, moved out of the client unchanged), server.toml with import, the CLI (init, run, start, stop, status, cmd, op, whitelist, backup, backups, restore, logs), crash restart, and the engine status endpoint.
2. install.sh and install.ps1, the Dockerfile, systemd and launchd units, service install, and the server archive in the release.
3. The web panel MVP: status, live TPS/MSPT/RAM/CPU, players, console, backups and login links.
4. horizon-bots and tools/server_bench.py, with a first run against vanilla (1/10/25/50 bots, mixed) published in benchmarks/: done, benchmarks/server-mixed-2026-10-18/, and with equal work benchmarks/server-equal-mixed-2026-10-18/ (section 9).
5. Small engine fixes found on the way: logging refused logins (vanilla's lines).

Where it stands (2026-10-16): items 1 to 4 are built and in review (section 11). What's left for the 23rd: run the Dockerfile and install.ps1 once on a machine that has Docker and on Windows; publish a release that holds horizon-server, so the installers have something to install; a clean benchmark run on a quiet machine, then published in benchmarks/; and item 5.

Out (next):

• the Nether and End for remote players;
• the headless engine and its RAM win;
• a server-only prepare without a JDK;
• a Windows service, hot reload, modpacks in the panel;

Online mode works, so a public server can check accounts. An offline server should still run only on a LAN, a tailnet, or behind a whitelist.

11. Pull requests

In order, each one small and shippable alone:

1. This design (docs only).
2. horizon-server, stacked in three parts:
   • a. The crate, the vanilla files moved out of the client unchanged (horizon-server-files), server.toml with import, init, run (crash restart, clean stop on Ctrl-C and SIGTERM), and op and whitelist for a stopped server: #551.
   • b. The control API on 127.0.0.1 with its token, start, stop, restart, status, cmd and logs, plus the engine status endpoint. That endpoint is about 30 additive lines in game/dedicated.rs plus a new game/dedicated/status.rs. perf-ticks-2's branches touch dedicated.rs elsewhere (listen: true), so the edits don't overlap. It also makes the dedicated loop record its tick times, so /tick query reports them: #552.
   • c. Backups, restore and automatic backups: #553.
3. Install scripts, Docker and services (service install), and horizon-server in the releases: #554.
4. Web panel MVP: #555.
5. Bots and the first benchmark against vanilla: #556.

Each PR is stacked on the one before. Merging them in order keeps every step shippable alone.