Roblox networking

Build explicit request and replication contracts in which the server decides authoritative game state and clients provide input or intent. Targets Roblox's rolling platform APIs. This skill goes deeper than the networking primer in roblox-luau.

When to use

• Use to design, implement, debug, or secure cross-boundary Roblox communication.
• Use when a remote trusts client values, an exploiter can target arbitrary Instances, messages spam services, streamed objects are missing, or clients disagree with the server.

When not to use: basic Luau/services belong to roblox-luau; persistent state belongs to roblox-datastores; physical ownership mechanics also compose with roblox-physics.

Workflow

1. Inspect the existing protocol. Find every remote and both endpoints; document direction, sender, payload, frequency, authority, validation, and consumers. Reuse the canonical remote folder—do not create a duplicate because discovery was skipped.
2. Classify each message. Client request, server fact, or ephemeral cosmetic sample. Choose reliable event, unreliable event, or request/response from semantics—not convenience.
3. Minimize the payload. Send stable identifiers and intent. Do not send a price, damage, ownership result, arbitrary path, or computed outcome the server can derive.
4. Validate in layers. Check type/shape/finiteness, allowlisted value, Instance class and ancestry, player permissions/state, distance/line of sight where relevant, server cooldown, and rate budget before doing expensive work.
5. Apply on the server. The server resolves targets and mutates health, inventory, currency, cooldowns, and progression. Client-side checks improve UX but grant no trust.
6. Replicate narrowly. Use FireClient for private or local facts; broadcast only shared facts. Avoid sending replicated properties again unless the client needs a distinct presentation event.
7. Handle time and lifecycle. Requests may arrive after death, respawn, streaming changes, or disconnect. Resolve the current character/state during handling and clean per-player limiter data.
8. Verify with Server & Clients. Exercise normal, malformed, spam, out-of-range, stale character, rapid respawn, leaving, simultaneous players, targeted, and broadcast cases. Inspect server and each client Output separately.

Choose the transport

Never invoke a client synchronously from the server. A client may disconnect, error, or never return. Prefer server RemoteEvent:FireClient() and a separate response event when needed.

Delivery rules that change handler design (per connection, per direction):

• RemoteEvent messages arrive in send order, even across different RemoteEvents, but order covers when each handler starts. A handler that yields (task.wait, WaitForChild, a DataStore call) lets the next message from the same player start before it resumes, so check-then-yield-then-apply code can approve two requests. Claim the cooldown or mark the request in progress before the first yield.
• A client firing a RemoteEvent faster than the throttle is delayed, not dropped; the backlog still arrives. UnreliableRemoteEvent traffic over the limit, over 1000 bytes, or with no handler connected is discarded.
• Messages sent before a handler connects are queued (bounded) for RemoteEvents and RemoteFunctions. There is no ordering guarantee between a RemoteFunction and a RemoteEvent; use one remote type or your own sequence number when order matters.

Pattern: validate before resolving gameplay

-- ServerScriptService/CombatRequests.server.luau
local Players = game:GetService("Players")
local ReplicatedStorage = game:GetService("ReplicatedStorage")
local Workspace = game:GetService("Workspace")

local attack = ReplicatedStorage.Remotes.Attack
local lastRequest: {[Player]: number} = {}
local RANGE = 12
local COOLDOWN = 0.25

attack.OnServerEvent:Connect(function(player: Player, target: unknown)
    local now = Workspace:GetServerTimeNow()
    if now - (lastRequest[player] or -math.huge) < COOLDOWN then return end
    lastRequest[player] = now

    if typeof(target) ~= "Instance" or not target:IsA("Model") then return end
    if not target:IsDescendantOf(Workspace.Characters) then return end
    local targetHumanoid = target:FindFirstChildOfClass("Humanoid")
    local targetRoot = target:FindFirstChild("HumanoidRootPart")
    local character = player.Character
    local root = character and character:FindFirstChild("HumanoidRootPart")
    local humanoid = character and character:FindFirstChildOfClass("Humanoid")
    if not targetHumanoid or not targetRoot or not root or not humanoid then return end
    if humanoid.Health <= 0 or targetHumanoid.Health <= 0 then return end
    if (root.Position - targetRoot.Position).Magnitude > RANGE then return end
    if not serverCombatStateAllowsAttack(player, now) then return end

    targetHumanoid:TakeDamage(serverDamageFor(player))
end)

Players.PlayerRemoving:Connect(function(player)
    lastRequest[player] = nil
end)

This is still only a compact example: a real melee system may require server-known attack windows, line-of-sight/shape checks, team rules, and lag policy. Do not treat one distance check as security.

Pattern: token bucket at the boundary

type Bucket = {tokens: number, updatedAt: number}
local buckets: {[Player]: Bucket} = {}
local CAPACITY, REFILL_PER_SECOND = 6, 3

local function consume(player: Player, cost: number): boolean
    local now = os.clock()
    local bucket = buckets[player] or {tokens = CAPACITY, updatedAt = now}
    bucket.tokens = math.min(CAPACITY,
        bucket.tokens + (now - bucket.updatedAt) * REFILL_PER_SECOND)
    bucket.updatedAt = now
    if bucket.tokens < cost then buckets[player] = bucket; return false end
    bucket.tokens -= cost
    buckets[player] = bucket
    return true
end

Assign cost by server impact. Reject cheaply before datastore calls, cloning, raycasts, or broad replication. Log aggregate abuse signals, not one warning per rejected packet.

Replication, streaming, and prediction

• Replicated Instances/properties are already a state channel. Use remotes for intent, private state, or presentation cues, not an unconditional parallel copy of the DataModel.
• With instance streaming, a valid server Instance may not exist on a client. Send a stable ID and tolerate absence; do not wait forever for optional streamed content.
• High-rate cosmetic data may use UnreliableRemoteEvent; make each sample self-contained because delivery and order are not guaranteed. Payloads over 1000 bytes are dropped (Studio Output reports the overage). Client-to-server traffic is throttled at roughly 500 calls/second per client for each remote type, shared by every RemoteEvent (and separately by every UnreliableRemoteEvent) in the experience — which is what a legitimate player hits before any attacker does.
• Without Server Authority, predict only latency-sensitive reversible presentation. Include a client sequence/command ID; the server returns authoritative state and acknowledgement; the client corrects smoothly. Never let prediction award damage, currency, inventory, or progression.
• Network ownership improves responsiveness but lets that client influence physical simulation. Validate gameplay consequences on the server; ownership is not authorization.

Places that use Server Authority

Check Workspace.AuthorityMode before designing prediction. When it is Server (opt-in, fully released 2026-07-09; setting it also enables next-generation replication, the Input Action System for player scripts, deferred signals, fixed simulation, and streaming), the engine supplies prediction, rollback, and resimulation, and gameplay objects can stay server-owned without input lag. Work with it instead of building a parallel scheme:

• Put core gameplay in a ModuleScript required by both a server Script and a client LocalScript, with logic in RunService:BindToSimulation() callbacks. Only properties and methods tagged Simulation Access are allowed there; use time() (synchronized and rewound during resimulation), not os.clock(), tick(), or os.time().
• Send simulation-affecting input through InputActions (their InputContext must sit under the owning Player), not UserInputService events. Keep synchronized custom state in Attributes on predicted instances (only the first 64 per instance replicate; names and string values up to 50 characters) and write them only inside bound simulation functions.
• Remotes still carry discrete facts and requests, but they are not ordered with property and attribute replication. Play effects and sounds from RenderStepped, reading simulation results, so a mispredicted frame does not trigger them twice.
• RunService:SetPredictionMode() forces prediction on or off per instance (client-only call). Test with Studio's Network Simulator to add latency, jitter and loss.

Server Authority is a project-wide migration. Do not switch AuthorityMode on an existing place as a side effect of another task; report it as an option instead.

Common failures

Resources

• Read references/validation-and-testing.md for payload rules, Instance/finiteness checks, replication design, and the required multi-client abuse matrix.

Related skills

• roblox-luau — execution locations and basic RemoteEvent mechanics.
• roblox-characters — respawn-safe character resolution.
• roblox-physics — network ownership, ray/overlap validation, and physical consequences.
• roblox-studio-workflow — Server & Clients testing and Output inspection.

Primary references

• https://create.roblox.com/docs/scripting/events/remote
• https://create.roblox.com/docs/scripting/security/client-server-boundary
• https://create.roblox.com/docs/projects/server-authority
• https://create.roblox.com/docs/physics/network-ownership
• https://create.roblox.com/docs/studio/testing-modes