mcpbeat

Godot Server Architecture

thedivergentai/godot-server-architecture

Expert blueprint for dedicated / headless multiplayer hosts: ENet/DTLS, authority validation, safe packet decode, matchmaker handoff, and health telemetry. Use when building authoritative servers, --headless hosts, or hardening host networking. Keywords: dedicated server, headless, ENet, DTLS, authority, safe_packet_decoder, multiplayer host, WebSocketMultiplayerPeer.

10k tokens
context cost
the whole folder, loaded on every use
16
files
ships runnable scripts
0
copies elsewhere
how many repositories repackaged it
451
stars on the repo
on the repository, not the skill itself

Install

one command, takes just this skill from the repository
npx skills add https://github.com/thedivergentai/GD-Agentic-Skills --skill godot-server-architecture

The instruction itself

30 sections, as written by the author

Godot 4.7 Baseline

  • Expert patterns in this skill target Godot 4.7+ (stable, 2026-06-18).
  • Consult the Godot 4.7 migration guide when upgrading projects from 4.6.
  • NEVER assume 4.6 defaults (stretch mode, audio area_mask, RichTextLabel percent flags) without checking 4.7 migration notes.

Server Architecture (Dedicated Host)

Authoritative multiplayer host patterns — headless boot, ENet/DTLS, packet safety, kicks, and telemetry.

> RID SceneTree-bypass rendering/physics swarms belong in godot-performance-optimization (and physics_server_direct.gd / rid_performance_server.gd only when the host sim truly needs server RIDs). Lobby/RPC gameplay APIs: godot-multiplayer-networking.

Skill boundary (Do NOT Load)

| Use this skill for | Use godot-multiplayer-networking for |

| :--- | :--- |

| --headless / dedicated export boot | Lobby UI, matchmaking UX, friend invites |

| ENet/DTLS host peer + safe decode | RPC signatures, @rpc gameplay handlers |

| Authority validation on privileged ops | MultiplayerSynchronizer / scene replication |

| Kick, health telemetry, matchmaker handoff | Client prediction, lag compensation |

Do NOT Load lobby/RPC tutorial scripts from multiplayer-networking when only booting a host — follow Host Golden Path here first.

Host Golden Path (MANDATORY)

  • Headless detect/initMANDATORY headless_init_manager.gd (--headless / dedicated_server feature).
  • Safe decodeMANDATORY safe_packet_decoder.gd before any untrusted get_var.
  • Host peer — enet_optimized_host.gd; add dtls_secure_server.gd when encrypting UDP.
  • Authority — server_authority_validator.gd on every privileged RPC.
  • Ops — peer_kick_manager.gd, server_health_exporter.gd; matchmaker handoff via server_matchmaker_client.gd.

Available Scripts

headless_init_manager.gd

Detect/initialize dedicated server logic for --headless / dedicated_server.

headless_manager.gd

Headless runtime manager companion patterns.

enet_optimized_host.gd

High-performance ENet UDP hosts with bandwidth/client limits.

dtls_secure_server.gd

DTLS + X509 hardening for ENet UDP.

safe_packet_decoder.gd

Forbid object decoding on untrusted packets (RCE guard).

manual_network_poll.gd

Manual multiplayer.poll() when auto-poll is disabled.

isolated_multiplayer_api.gd

Isolated MultiplayerAPI instances (client+server in one process).

server_authority_validator.gd

get_remote_sender_id() gates for authoritative requests.

websocket_server_compat.gd

HTML5-compatible WebSocketMultiplayerPeer hosts.

peer_kick_manager.gd

Graceful peer termination with reason propagation.

server_matchmaker_client.gd

Load-balancer / matchmaker handoff to game hosts.

server_health_exporter.gd

Headless telemetry for monitoring stacks.

physics_server_direct.gd / rid_performance_server.gd

Optional host-side RID sim — only when node physics cannot hold tick budget: > ~200 active bodies per tick, or headless host CPU > 70% on physics step with nodes. Criteria: profile first; if SceneTree bodies dominate, prefer godot-performance-optimization. Do NOT Load RID scripts for ≤64 entity lobbies.

NEVER Do in Server Architecture (Host)

  • NEVER trust the client — Validate state, purchases, and damage on the authoritative host.
  • NEVER use TRANSFER_MODE_RELIABLE for continuous streams — Prefer unreliable for high-rate transforms.
  • NEVER use get_var(true) on untrusted packets — Object decode = RCE. MANDATORY safe_packet_decoder.
  • NEVER use TCP for fast-paced action — Prefer ENet UDP (or WebSocket for HTML5 constraints).
  • NEVER run a dedicated server without stripping visuals — Dedicated Server export / dummy drivers.
  • NEVER expect RPCs before connected_to_server / peer ready.
  • NEVER assume UNRELIABLE packets arrive in order.
  • NEVER leave SceneTree.multiplayer_poll false without manual poll().
  • NEVER mix incompatible engine/multiplayer protocol versions across peers.
  • NEVER forget free_rid on server-created RIDs if the host uses Physics/RenderingServer pools.

Host Patterns

Interest management

Large worlds: MultiplayerSynchronizer.public_visibility = false + visibility filters (AABB / grid) so the host does not sync the entire world to every peer.

# Hook on synchronizer — filter peers by grid cell / AABB (no full tutorial)
func _visibility_filter(for_peer: int, node: Node) -> bool:
	return _interest_grid.is_visible_to_peer(for_peer, node.global_position)
# Assign: synchronizer.set_visibility_filter(_visibility_filter)

Health metrics

Watch host FPS, static memory (RID leaks), and orphan counts via server_health_exporter.gd.

Deep recipes (on demand)

> LLM-ignorance rule: if a general agent would not know it before reading, it lives here or in scripts/ — never delete, only move.

| Topic | Reference |

|-------|-----------|

| RID canvas/physics cookbook | rendering-physics-server-cookbook.md |

Reference

> Progressive disclosure: open Official Documentation links only when researching a specific API; load Related Skills when routing to a peer domain — do not preload the whole lattice.

Official Documentation

  • Using Servers — RID-based RenderingServer/PhysicsServer/NavigationServer workflow when SceneTree nodes are too slow.
  • RenderingServercanvas_item_* / instance_* / free_rid for procedural draw and mesh swarms without MeshInstance nodes.
  • PhysicsServer3Dbody_create, space binding, and direct-state queries for headless authoritative simulation.
  • PhysicsServer2D — 2D body/shape RIDs mirroring the same SceneTree-bypass pattern.
  • RID — opaque server handles; every *_create() needs a matching free_rid to avoid leaks.
  • High-level multiplayer — authority, RPCs, and peer lifecycle for dedicated hosts and isolated MultiplayerAPI branches.
  • ENetMultiplayerPeer — UDP host creation, channels/bandwidth limits, and DTLS host setup on peer.host.
  • WebSocket multiplayer — browser-compatible peer path when ENet UDP is unavailable (HTML5 clients).
  • Exporting for dedicated servers — dedicated-server export presets and stripping visuals/audio for production hosts.
  • Command line tutorial--headless and CLI flags used by headless init/managers.
  • Binary serialization APIget_var(false) / object-decoding rules that block RCE on untrusted packets.
  • DTLSServer — DTLS accept path complementary to ENet dtls_server_setup with X509/TLSOptions.
Prerequisites
  • godot-project-foundations — project layout, Autoloads, and feature tags that dedicated-server and headless launches depend on.
  • godot-gdscript-mastery — typed RID arrays, @rpc annotations, and safe Variant decoding patterns used across server scripts.
  • godot-physics-3d — node-level PhysicsBody3D/space concepts before bypassing them with PhysicsServer3D RIDs.
Complements
Downstream / consumers
Master
  • godot-master — library router and mirrored module entry; open when discovering which Domain Skill owns a cross-cutting server concern.

How to use it

Copy the folder

Take thedivergentai/godot-server-architecture from the repository into ~/.claude/skills for personal use, or into .claude/skills inside a project.

Check the name does not clash

The agent identifies a skill by the name field in its header. Two skills with the same name cannot sit side by side — one of them will be ignored.