Best Minecraft Mods for Playing With Friends

Nearly every "best Minecraft mods" list is written for one person playing alone, which is why so much of that advice falls apart the moment a world is shared. Half the mods people recommend for lag do nothing at all on a server. A few of them stop a server booting. And the mods that matter most to a group, claims, voice chat and shared maps, rarely appear on those lists because a solo player never needs them.

This guide is organised around the only question that decides whether a mod works for a group: which machine does it go on. Every mod below carries the side it actually needs, read off the mod authors' own environment tags rather than from memory, and checked against 26.2, the current Minecraft release.

A Minecraft mod is a jar file, and a jar file only does something on the machine it is installed on. That sounds obvious until you notice how much advice ignores it. Modrinth publishes an environment tag on every project, one value for the client and one for the server, and the four combinations cover every mod in this guide.

Read the four shapes once and most multiplayer mod problems stop being mysterious:

• Client required, server unsupported. A pure client mod. Nothing happens if you put it on the server, and on Forge or NeoForge it will usually stop the server booting. Sodium, Iris, Entity Culling, Mouse Tweaks, Inventory Profiles Next and Controlling all live here. Nobody in the group has to match anybody else.
• Client unsupported, server required. A pure server mod. Nothing happens if you put it in your own mods folder. Alternate Current, ServerCore, Flan and every Bukkit plugin are in this group. One person installs it once and the whole group gets the benefit.
• Client required, server required. A content mod. It adds real blocks, items or entities, so the server needs it to know they exist and every client needs it to draw them. Versions must match or the connection is rejected outright. Sophisticated Backpacks, Waystones and Cobblemon are here.
• Optional on both sides. The confusing one, and the source of nearly every support ticket. It does not mean "install it wherever". It means each side works without the other, and the two installs usually do completely different jobs.

This is the table to bookmark. Side is what the mod actually needs, not what a wiki page says: where a mod author has contradicted or clarified their own tags, the author wins. Loaders lists what is available on current Minecraft, because a mod that has no build for your loader is not a recommendation, it is a dead end.

Either side means the mod genuinely works installed on one side without the other, and helps whichever side it is on. Lithium is the clearest example, and its own project page says so directly: it "can be installed on servers without requiring clients to also have the mod (and vice versa)".

The practical split is simple. Client only mods are everyone's own business. Server only and Both mods are a group decision, and Both mods are the ones that will kick someone at the login screen if a single version is wrong.

Two map mods explain this better than any definition. JourneyMap is optional on both sides, and its own page spells out what each install does: the client copy is what draws the map, and the server copy draws nothing at all, existing only so an admin can switch off cave mapping or the entity radar for everyone. Install it only on the server and no player sees a map.

Xaero's Minimap is a different shape again: client required, server optional. The client copy is the mod. The optional server copy buys reliable multi-world detection, so waypoints stop bleeding between dimensions, plus rank-based and potion-effect restrictions for anyone running PvP.

There is a second trap in the tags. Some server-only mods look like they work in singleplayer, and they do, because singleplayer runs an integrated server inside your client. ServerCore says so in its own readme while Modrinth marks it client unsupported, and both are true. It works in your own world and it still never belongs in your mods folder when you join a friend.

Here is the mistake this guide exists to fix. The three most recommended performance mods on the internet, Sodium, Iris and Entity Culling, are all marked server unsupported. They are renderers. They change what your graphics card draws, and a dedicated server draws nothing. Entity Culling puts it in one line in its own FAQ: "This is a fully client-side mod and does not need to be installed on the server." If your problem is that blocks reappear after you break them and mobs stutter in place for everyone at once, no client mod will touch it. Our guide to Minecraft performance mods covers the frame rate half of this properly, so what follows is only the part that changes a shared world.

Drop a client renderer into a server mods folder and you do not get a small improvement, you get a crash. The server jar has no client rendering classes, so Forge and NeoForge die at startup complaining about loading a client class for an invalid dedicated server dist, and Fabric throws a NoClassDefFoundError instead.

Every mod here goes in the server mods folder. Nobody in the group installs anything.

1. Lithium first, always. It rewrites physics, mob AI, block ticking and fluid flow with the stated goal of changing no vanilla mechanic, and it is the largest single reduction in tick time available to a small server. It needs nothing from the client.
2. FerriteCore next, because it is free. It deduplicates blockstate and model data in memory, and the project cites a large Fabric pack falling from 1,792 MB to 984 MB of heap. On a memory-tight plan that is the cheapest gigabyte you will ever recover.
3. Chunky before you invite anyone. It pre-generates a radius of chunks in bulk, which removes the single most common source of hitching: someone sprinting into terrain that has never existed before while everyone else waits for the tick to catch up.
4. C2ME if people explore constantly. It parallelises chunk generation, loading and I/O across cores. It is marked server required, and the client optional tag exists only because singleplayer runs an integrated server. It is still alpha versioned, and Fabric and NeoForge are two separate projects rather than one download with two files.
5. ScalableLux alongside C2ME rather than instead of it. It is the maintained fork of Starlight, and its readme frames the payoff exactly right: vanilla lighting is a bottleneck for high-performance chunk generation, and that matters most on dedicated servers with enough players to stress it.
6. Alternate Current the moment anyone builds redstone. It replaces the power calculation for redstone dust, and its page puts the MSPT contribution of dust at up to 30 times lower. Client unsupported, server required, and a perfect illustration of the whole point: a redstone-heavy world lags for everybody at once and no client mod can help.
7. ServerCore if you still need more, with care. It gives you entity activation ranges, mob caps, dynamic view and simulation distance and chunk-tick throttling. Its own readme is honest that the features which change vanilla behaviour are off by default, so the aggressive settings are opt-in and your players will notice them.

Krypton is tagged optional on both sides, but every benefit in its readme is phrased as server CPU and server tick time, which makes it a server mod that happens to be harmless on a client. Two things to know: it is Fabric only, with no Forge or NeoForge port, and the author opens the readme by saying it is a work in progress with no stability guarantees.

VMP is server required and targets high player counts specifically: chunk sending, packet priority, per-player render distance and entity tracker caching. For four friends on a survival world it is the wrong tool and Lithium is the right one. It is Fabric only, still shipping on Modrinth's alpha channel, and its readme warns that things are expected to break and recommends running it alongside Lithium.

Moonrise deserves a note of its own because the name is a trap. It optimises the integrated and dedicated server, absorbing the old Starlight light engine rewrite plus chunk system and collision work, and its page says plainly that it can run on either side but primarily benefits the server. The Modrinth project is moonrise-opt. The bare slug moonrise is a completely unrelated 2023 mod that moves the moon across the sky, client only, and stuck on MC 1.20.4.

All three still appear on lists published this year. None of them is installable on current Minecraft in the way those lists imply.

• Starlight: archived on Modrinth, last release December 2023, tops out at MC 1.20.4. Its own page points at Moonrise. Mention it as the ancestor of Moonrise and ScalableLux, not as an option.
• Noisium: delisted from Modrinth search, GitHub repo archived, last release June 2025 for MC 1.21.6. If you came looking for worldgen performance, C2ME is the live answer.
• ModernFix: still active, but there is no 26.2 build as of publication, and on the 26.x line it is NeoForge only. Its last Fabric builds were in late 2025 for MC 1.20.1 and 1.21.1, so a Fabric group cannot use it at all. Its wins are also mostly launch time and client memory rather than tick rate.

When someone says the server is lagging and someone else says their game is fine, install Spark on both. It is a profiler, available as a mod on every loader and as a plugin for Paper, Spigot and Velocity, and it fixes nothing, which is exactly why it is useful.

Run the profiler on the server and you get a call tree showing which mod is eating tick time. Run it on the client and you get frame time attribution. Within five minutes you know whether you need the first half of this section or the second, with evidence instead of a group argument.

This is the half of the mod list that gets ignored in server guides and it is the half your friends will actually notice. The client-only entries need no coordination at all, which makes them the easiest wins available: one person can install them, or all five, and nothing about the server changes.

JEI is still tagged optional on both sides, and on current Minecraft that tag is actively misleading. Since Minecraft 1.21.2 recipes are stored on the server rather than the client, and JEI's own FAQ now says: "Please install JEI and NeoForge/Fabric on the server to sync recipes to the client."

If you have ever installed JEI, joined a friend and found an empty or half-broken recipe list, that is the cause. Treat JEI as a both-sides mod on anything from 1.21.2 onward. The same server-side recipe change applies to REI, so install that on the server too. REI even has a dedicated Server Component project, client unsupported and server required, whose only job is serving recipe data, and the REI page suggests running JEI alongside it if you want to support users of both.

EMI is the nicest of the three to use and the one you probably cannot have. Its newest release is for MC 1.21.1, confirmed identically on Modrinth, CurseForge and GitHub, with nothing for 1.21.2 or any 26.x version. The repository is still active, so this is a porting backlog rather than an abandoned project, but recommend it only to a group deliberately playing an older pack.

A minimap is per player by design. Your explored map data and your waypoints live on your machine, which is why your friend's carefully labelled waypoint list never shows up for you, and why nobody in the group ever has the same picture of the world.

The fix for a group is BlueMap on the server. It is client unsupported and server required, it renders the world into a 3D map served over a web page, and it runs as either a mod or a plugin from the same project. Nobody installs anything: you send them a link. For a friend group it is the lowest-friction mod in this entire guide, and it pairs well with pre-generating the world so the map is complete before anyone walks out there.

For personal navigation on top of that, Xaero's Minimap and Xaero's World Map are deliberately separate downloads so people who only want the corner map are not carrying the fullscreen one. Both are client required with an optional server copy. JourneyMap is the alternative if you want the browser view and the in-game map from one project, and it has a Paper build for servers that are not modded at all.

Jade and WTHIT both tell you what block or entity you are looking at, and both draw the same line in their own words: basic identification works client-only, but anything backed by server-held state needs the server copy. Jade lists item storage, brewing stand fuel and bees in a beehive. WTHIT says anything reading NBT data, such as inventory contents, probably needs it on the server. Jade has far more users, WTHIT has the wider loader coverage on 26.2.

AppleSkin sits in the same middle ground. Installed only on your client it shows food values fine, but the saturation and exhaustion numbers are state your client cannot know on its own, so those need the server copy too. The author states this on the project page rather than leaving you to discover it.

The pure client-side group needs no discussion at all, which is the point: Inventory Profiles Next for sorting, locked slots and gear sets, Mouse Tweaks for drag and scroll-wheel item movement, and Controlling for a search bar in a key bindings screen that a 150-mod pack has turned into an unreadable wall. All three are client required and server unsupported. Install them or do not, nobody else is affected.

Sophisticated Backpacks and Sophisticated Storage are the opposite case and the cleanest teaching example in this guide. They are required on both sides at matching versions, because they place real items and blocks in the world: the server has to know those exist and every client has to draw them. A version mismatch is a rejected connection, not a degraded experience. They are also NeoForge only on 26.2, with no Fabric build at all, which is a hard stop for a Fabric group.

Most claim guides are written for public servers full of strangers. A friend group has a different problem: not malice, but the friend who mines through your wall to get home, the accidental lava, and the question of who took the diamonds out of the shared chest.

The first decision is not which claim system but which server software you are on, because mods and plugins are not the same thing and cannot run together.

Three real options, in the order worth trying them.

• Open Parties and Claims is the default recommendation. Server required, client optional, current on 26.2, and available on all four loaders. The client copy only adds the claims UI, keybinds and an overlay on the Xaero maps, so a friend with a vanilla client can still join. Its config scales limits with headcount through claim and force-load bonuses per party member, which is the behaviour you want when the group grows.
• Flan is the pick when you want zero client installs and claim regions that are not chunk shaped. Its readme is blunt about it: Flan is completely server side only, so you can set up a Fabric server with Flan and connect with a vanilla client and it works just as fine. No 26.2 build as of publication.
• FTB Chunks is what you already have if you are running an FTB or CurseForge pack, because most of them ship it. It is genuinely both sides, since it contains a minimap, and its claims belong to an FTB Team rather than to a player, which is the mechanic that makes it work for a group. Its ceiling today is 26.1.2.

GriefPrevention, CoreProtect, WorldGuard and Towny are Bukkit-ecosystem plugins. They need Paper, Spigot, Purpur or Folia and will not load on Fabric, Forge or NeoForge, and none of the claim mods above will load on Paper. If you have not installed one before, how to install a Minecraft plugin is the ten-second version.

CoreProtect is the one to install first, whatever else you choose. It does not prevent anything, it logs every block and container change so you can roll it back and see who did it. For a friend group that is worth more than a claim system, because the actual questions are "who flooded the basement" and "where did the diamonds go", and only a logger answers those. We have a walkthrough for installing CoreProtect.

GriefPrevention is the self-service claim option: players claim land with a golden shovel and containers, doors and blocks inside are protected automatically, with no admin involvement. WorldGuard is the opposite model, where an admin draws regions and sets flags, which is usually overkill for five people except for protecting a shared spawn. It needs WorldEdit, covered in using WorldEdit on your server.

Towny is excellent and probably not for you. It is built for public servers with dozens of towns and strangers who need taxes and residency rules. Five friends who trust each other get bureaucracy instead of protection. If you want per-rank claim limits on any of these, LuckPerms is what the claim systems read those limits from.

One version note for accuracy: GriefPrevention and CoreProtect both top out at 26.1.2 as of publication. CoreProtect published a build in July 2026 and GriefPrevention in March 2026, so both are active and have simply not tagged 26.2 support yet.

Claim mods usually also do chunk force-loading, and force-loading is the part that costs you tick time. A force-loaded chunk keeps ticking whether or not anyone is near it, which is exactly what a player wants for their farm and exactly what your server does not want multiplied by five friends.

FTB Chunks' own defaults are worth reading before you hand them out: 500 claimed chunks and 25 force-loaded chunks per player, no hard team limit on either, and a party limit mode that can be set to sum the allowance of every member. On the sum setting, five friends in one team hold 125 permanently force-loaded chunks between them. A modpack can override any of this in its config, so check what your pack actually shipped rather than assuming.

Two protection options in the same config are disabled by default for performance reasons the config comments state outright: flowing fluid protection and fire spread protection both hook checks that run very frequently. Turn them on deliberately, not by reflex.

Proximity voice changes a shared world more than any other mod on this page, because it turns a server from a place you build in into a place you hang out in. Two projects are worth considering, and the difference between them is almost entirely operational.

The side is the same for both: it goes on the server and on the client of anyone who wants to talk. Neither forces the issue, though. A player without the mod still joins and plays normally and simply has no voice, which Simple Voice Chat states in its FAQ and Plasmo Voice states in its docs.

Voice runs over UDP and the game runs over TCP. Opening the TCP game port does not open UDP, and this is the single most common reason a voice chat install appears to do nothing.

The two projects default to opposite answers. Simple Voice Chat uses port 24454 by default, set in voicechat-server.properties, and its wiki explicitly warns against setting it to reuse the Minecraft port, because UDP on that port is also used for the server query and doing so may crash the server. Plasmo Voice defaults its port to 0, which its docs state means it uses the same port as your Minecraft server: game on 25565 TCP, voice on 25565 UDP.

On a hosted server this is a panel setting rather than a router argument. Our walkthrough for setting up Simple Voice Chat covers the port and the config on a CraftLands server. Some packs ship it already, including Fear Nightfall, where the group talks positionally from the first night.

Two more things worth knowing before you promise the group voice. Simple Voice Chat only needs the first two numbers of the client and server versions to match, so 1.2.0 and 1.2.5 talk to each other while 1.1.x and 1.3.x do not. And it is Java only: the author states that a voice chat is impossible to implement in Bedrock Edition for technical reasons, so anyone joining through Geyser gets no voice at all.

A content mod scales well for a group when one person doing the work benefits everybody, or when several people can work on the same thing without getting in each other's way. Most content mods fail that test quietly: they are fine in multiplayer but nothing about them is better because your friends are there.

These are the ones where the multiplayer claim comes from the mod itself rather than from a reviewer.

• Waystones (both sides, has a 26.2 build). It positions itself as a multiplayer alternative to /warp and /home, and its feature list includes global waystones for community hubs and sharing with friends. One player activates a waystone and everyone can use it, which is the cleanest example of the pattern on this page.
• Lightman's Currency (both sides, MC 1.21.1 and 1.20.1). Its own one-line description is a currency mod that lets you safely trade in a multiplayer environment. The part that matters for a group in different time zones: trading machines are protected so only the owner can break them, so a shop keeps selling while you are asleep.
• BlueMap (server only, current on 26.2). Nobody installs anything, everybody sees the same world. Covered above under maps, but it belongs on this list too, because a shared picture of the map changes how a group plans.
• Cobblemon (both sides). Its page states the multiplayer case directly: you can trade and battle with your friends. It is the most social loop in modded Minecraft and needs nobody to learn a machine. Note the version cost, though: it is pinned to MC 1.21.1, so choosing it means running a 1.21.1 server.
• MineColonies (both sides, actively updated on 1.21.1). Each player runs their own colony of AI citizens in the shared world, which gives a group parallel projects instead of one shared to-do list. Worth knowing that colonies tick constantly for citizen pathfinding, so several colonies cost the server more than several players, and the mod exposes colony size and distance configs for exactly that reason. Life in the Village 3 is built around it if you want the pack rather than the mod.
• Create (server required, and every client needs it too, whatever the client optional tag suggests). The co-op factory mod. One person builds the power, another the processing line, and the contraption is a shared object rather than a personal one. Two honest caveats: it stops at MC 1.21.1, and moving contraptions and trains are simulated on both sides and reconciled every tick, so it is more latency sensitive than a static machine.

Terralith and Tectonic are both server required with the client optional, because the terrain is authored on the server: a client without either mod still sees the world exactly as the server generated it. Biomes O' Plenty is required on both sides instead, because it registers real new blocks rather than rearranging vanilla ones. That contrast is the whole client and server distinction in one category.

What matters more is timing. Terralith's own page states it in capitals: you cannot remove it from an existing world, and adding it to one is not recommended. The mechanism is simple. Already-generated chunks keep vanilla terrain, new chunks use the new noise settings, and the two meet as a vertical wall along a chunk border. Remove it later and the save references biome IDs the server no longer has. Settle your worldgen mods before the first chunk generates, take a backup before you touch anything, and pre-generate afterwards rather than before.

Five mechanisms cover almost every "it worked in singleplayer" report. Knowing which one you are looking at turns an evening of guessing into a two-minute fix.

1. A client renderer in the server mods folder. Sodium, Iris and Entity Culling are all server unsupported. On Forge and NeoForge the server dies at boot complaining about loading a client class for an invalid dedicated server dist, and on Fabric you get a NoClassDefFoundError. Shaders, minimaps, zoom mods and FPS mods are client business, always.
2. Per-player world state. Distant Horizons is the clearest case and documents it itself: with the mod only on your client, the distant terrain is whatever you personally explored and saved, so your friend builds a castle and you keep seeing the empty hill from three weeks ago. Install it on the server as well, or use its separate Bukkit plugin, and the server sends level-of-detail data to every player who has the mod. Minimap waypoints have the same shape with lower stakes.
3. Desync under latency. Any mod that simulates continuous motion on both sides and reconciles each tick will diverge when packets are late, because the client is extrapolating. Valkyrien Skies 2 has an open issue of exactly this shape: dismount a ship helm while the ship is doing at least 6 m/s and you get thrown out of the ship, reproducible about 60 percent of the time with only that mod and its addons installed. Moving platforms, trains and ships are where you meet this class of bug.
4. Worldgen changed after the world exists. Covered above. This one is not a crash, it is a permanent scar on the map and a save that cannot cleanly go back.
5. Versions that do not match exactly. Not just the mods. NeoForge tells you outright to make sure you are using the same mod and NeoForge versions as the server, so the loader version counts too. Simple Voice Chat is looser and still refuses a client whose first two version numbers differ from the server build.

The error you get depends on your loader, and this is the part people get wrong when they search for a fix.

NeoForge and Forge check the mod list and network channels during the connection handshake and refuse with a dedicated screen. You get "Server mod list is not compatible" or "Server network message list is not compatible", followed by a table telling you exactly which mods your client is missing, which mods to remove, or which versions do not match. It is genuinely helpful.

Fabric Loader performs no mod list handshake at all, so a mismatched client is usually allowed to connect and then fails later, during registry sync, with a Fabric API message about receiving registry entries that are unknown to this client, and a note that this is usually caused by a mismatched mod set between the client and server. It names the namespaces rather than the mods, so you get a hint instead of an answer.

And a vanilla client hitting a modded server that uses neither system typically just sees a generic connection lost with an internal netty decoder exception. Plugins produce none of these, because plugins add nothing to the client at all.

Everything above reduces to one operational rule. Every player needs the identical loader, the identical loader version and the identical mod list, and the server needs to match them. Sending five friends a list of eight Modrinth links and hoping is how you spend the first evening reading crash logs instead of playing.

A modpack is nothing more mysterious than a pinned mod list. It fixes every mod version and the loader version in one file, which is why distributing a pack export from CurseForge or Modrinth, or a shared Prism instance, is a completely different experience from distributing instructions. Our guide to the best Minecraft modpacks covers 33 packs with the loader, the Minecraft version and the server RAM each one needs, and the full modpack library has a setup page for each of them.

On the server side the same rule applies with one extra step: the loader has to be installed before the mods, at the matching version. On CraftLands the pack and its loader go on before first boot and update from the panel when a new version ships, so there is no window where the server is running the wrong thing. Any pack outside the featured list installs the same way, covered in installing a custom modpack, and an existing server switches packs from the panel using installing a modpack.

Do these in order and the group starts playing on the first evening rather than the third.

1. Agree the loader and the Minecraft version first, because that decision silently removes options. Fabric rules out Sophisticated Backpacks and ModernFix. NeoForge rules out Krypton and VMP. Paper rules out every mod here except the handful that also ship as plugins, Chunky, BlueMap, Spark and both voice chats, and gives you every plugin instead.
2. Pin the client list as a pack export and send one file. Never a list of links.
3. Install the server-side set on the server: Lithium, FerriteCore, Chunky, a claim system, and BlueMap if you want the shared map. None of them needs a matching install on anyone else.
4. Pre-generate a few thousand blocks around spawn before anyone joins. It is the single most effective thing on this list and it only has to be done once.
5. Add voice chat and confirm the UDP port is open before the first session, not during it.
6. Turn on automatic backups now, while there is nothing worth losing.

RAM on a modded server is a fixed cost for the pack plus a roughly linear cost per concurrent player, which is the formula our Minecraft RAM calculator uses:

Round the result up to the next plan tier: 4, 6, 8, 10, 12, 16 or 24 GB. As rough floors for around 20 concurrent players, vanilla Java and Paper want 6 GB, Fabric 8 GB, and a Forge or NeoForge modpack 12 GB. Pricing is per gigabyte, so the RAM figure is the price: at €2.50 per GB per month a 4 GB server is €10 and an 8 GB server is €20, with the full tiers on the pricing page.

Every CraftLands Minecraft server runs on the AMD Ryzen 9 9950X with NVMe storage, which matters more than RAM once the server is ticking a claim system, a voice chat and a hundred mods, because Minecraft tick processing is largely single-threaded and bound by per-core speed. DDoS protection and automatic backups are included, locations are worldwide so you can pick the one closest to most of the group, and deployment takes under a minute. If you would rather build one yourself first, how to make a Minecraft server walks through the manual route including Forge and Fabric.

Start with the table. Put the server mods on the server, let everyone install their own client mods, and only coordinate the ones marked Both. That single habit removes most of the problems people spend evenings debugging.

When you are ready to give the group somewhere permanent to play, a CraftLands Minecraft server starts at €10 a month, deploys in under a minute, and installs your modpack with its loader before first boot so every player is matched to it from the first login. If you are unsure what size you need, the RAM calculator answers it in about thirty seconds.