A Rust toolkit for building a complete Guild Wars game database.

Vision

TyriaExtractor is a collection of extraction and reverse-engineering tools for turning Guild Wars client data into a comprehensive, machine-readable database. The goal extends beyond skills and items to monsters, NPCs, quests, and other game resources useful to players, community applications, build planners, archival projects, and technical research.

Game data comes from the user's official local client:

• static resources, executable tables, localized text, models, and textures from Gw.dat or Gw.snapshot;
• relationships supplied only at runtime, such as the model_id -> model_file_id item mapping, from official server-to-client messages observed by the injected sniffer;
• no wiki, website, third-party database, or remote API is used as a data source.

Quick start (Windows)

Install the 32-bit MSVC target, then build the native extractor and the 32-bit injector/sniffer pair:

rustup target add i686-pc-windows-msvc
cargo build --release -p tyria-extractor-rs
cargo build --release --target i686-pc-windows-msvc -p tyria_injector -p tyria_sniffer

After the official client has logged in and loaded a map, inject the DLL:

.\target\i686-pc-windows-msvc\release\tyria_injector.exe Gw.exe .\target\i686-pc-windows-msvc\release\tyria_sniffer.dll

The sniffer writes immutable, session-scoped evidence under captures/<session-id>/, with one JSONL per record contract: tyria_items.jsonl, tyria_quests.jsonl, tyria_npcs.jsonl, tyria_vendor_context.jsonl, tyria_collectors.jsonl, tyria_merchants.jsonl, tyria_crafters.jsonl, and tyria_skill_trainers.jsonl, plus the shared metadata sidecar tyria_capture.jsonl. Repeat --packet-log to merge the streams and capture sessions required by an extractor:

cargo run --release -- extract skills --snapshot "C:\path\to\Gw.dat"
cargo run --release -- extract images --snapshot "C:\path\to\Gw.dat"
cargo run --release -- extract items --snapshot "C:\path\to\Gw.dat" --packet-log ".\captures\<session-id>\tyria_items.jsonl"
cargo run --release -- extract quests --snapshot "C:\path\to\Gw.dat" --packet-log ".\captures\<session-id>\tyria_npcs.jsonl" --packet-log ".\captures\<session-id>\tyria_quests.jsonl" --item-log ".\captures\<session-id>\tyria_items.jsonl"
cargo run --release -- extract npcs --snapshot "C:\path\to\Gw.dat" --packet-log ".\captures\<session-id>\tyria_npcs.jsonl" --packet-log ".\captures\<session-id>\tyria_collectors.jsonl"
cargo run --release -- extract vendors `
  --snapshot "C:\path\to\Gw.dat" `
  --packet-log ".\captures\<session-id>\tyria_npcs.jsonl" `
  --packet-log ".\captures\<session-id>\tyria_vendor_context.jsonl" `
  --packet-log ".\captures\<session-id>\tyria_collectors.jsonl" `
  --packet-log ".\captures\<session-id>\tyria_merchants.jsonl" `
  --packet-log ".\captures\<session-id>\tyria_crafters.jsonl" `
  --packet-log ".\captures\<session-id>\tyria_skill_trainers.jsonl"

For recurring work, the Bun scripts select the newest numeric captures/<session-id>/ directory automatically and use C:\Program Files (x86)\Guild Wars\Gw.dat by default:

bun run build:capture
bun run inject
bun run regen
bun run extract:vendors

Set GW_DAT to use another archive, or CAPTURE_DIR to select a capture path or session ID instead of the newest one. Run bun run to list every target.

All extraction commands write below output/ by default. --out-dir <PATH> overrides that common parent. Generated output remains local and Git-ignored:

output/
  items/          items.json, capture.json, model_file/*.png
  skills/         skills.json, model_file/*.png, model_file_hd/*.png
  images/         manifest.json, png/*.png
  quests/         quests.json, capture.json
  npcs/           npcs.json, capture.json
  vendors/
    collectors/   collectors.json
    merchants/    merchants.json
    crafters/     crafters.json
    skill_trainers/ skill_trainers.json
    coverage.json
    capture.json

Capture-backed extraction requires format version 5 and rejects missing health metadata, packet loss, write failures, incompatible world-packet schemas, duplicate or missing capture_seq values, and every pre-cutover JSONL.

Extraction status

Legend: ✅ Done · 🚧 In progress · ⬜ Not started

Dataset	Status	Current result
Skill template corpus	✅	1,488 distinct IDs: 1,333 non-PvP skills/special rows plus 155 linked PvP variants, with metadata and text in all 11 client languages
Skill icons	✅	Standard- and high-resolution PNG icons
DAT textures	✅	62,314 web-ready PNGs from every decodable ATEX, ATTX, DDS, and inline FFNA texture in the current archive
Items	🚧	377 observed identities in the current capture; all names and 351 descriptions resolve in all 11 client languages
Item icons	🚧	Verified linked stream-1 PNGs for 309 of 333 observed model_file_id values; unresolved mappings are not guessed
Quests	🚧	11 observed active quests with localized fields, objective variants, rewards, and NPC role/map evidence
Monsters / creatures	⬜	—
NPCs	🚧	99 observed models across maps 146, 148, and 194, with runtime metadata and names in all 11 client languages
Vendors and skill trainers	🚧	2 collectors / 2 offers, 4 merchants / 48 items, 0 crafters, and 1 trainer / 112 skills in the current capture corpus

Items remain in progress because the decoding and export pipeline works, but the official client must still receive every relevant runtime item definition before the catalog can be considered exhaustive.

Quests remain in progress because no complete static quest table has been confirmed. The consumer preserves relations observed in official-client packet captures, but exhaustive quest, branch, dialogue, and prerequisite coverage is not yet proven.

NPC extraction consumes the dedicated NPC properties, spawn, and map-context stream. Coverage is observational: the present corpus spans maps 146, 148, and 194 and is not an exhaustive NPC list.

Vendor extraction emits stable service-instance/item/skill joins in output/vendors/{collectors,merchants,crafters,skill_trainers}/, with one same-named JSON file per directory. A service instance is identified by (map_id, npc_model_id, position) because multiple NPCs can share one model. output/vendors/coverage.json lists those observed instances by map; it is a traversal progress ledger, not a claim that every outpost has been covered. Service entries use the per-agent AGENT_UPDATE_NPC_NAME EncString when their capture observed it and resolve it from Gw.dat in all 11 client languages.

Data flow

Gw.dat / Gw.snapshot
  -> MFT and hash lookup
  -> archive decompression
  -> executable tables, localized text records, models, and textures

official client + injected sniffer
  -> tyria_items.jsonl: ItemGeneral runtime pairs and item EncStrings
  -> tyria_quests.jsonl: quest packets and dialogue relations
  -> tyria_npcs.jsonl: NPC properties, per-agent names, spawns, model/file relations, and maps
  -> tyria_vendor_context.jsonl: merchant-window and owner packets
  -> tyria_collectors.jsonl: collector exchanges
  -> tyria_merchants.jsonl: merchant item lists
  -> tyria_crafters.jsonl: crafter product lists
  -> tyria_skill_trainers.jsonl: skill-trainer lists
  -> local Gw.dat text and image resolution
  -> output/{items, skills, images, quests, npcs, vendors}/

The DAT remains the source of localized text records and assets. Runtime capture provides joins that are not available as a confirmed complete static table, most notably model_id -> model_file_id, plus each runtime item's name and info_string EncStrings. Active quests are queried once after hook installation so their QUEST_DESCRIPTION records are captured even when the client already knew the quests before injection.

Each data JSONL contains only its own packet or runtime-hook records. The shared world_packet_schema, hook-status, loss, and write-counter metadata lives once in the sibling tyria_capture.jsonl sidecar, which extraction discovers automatically. A monotonic capture_seq preserves cross-file event order when consumers merge the required streams. Only capture format 5 with its tyria_capture.jsonl sidecar is accepted. The standard tyria_items.jsonl always contains complete ItemGeneral records, decoder observations, and runtime item strings. TYRIA_VERBOSE_JSONL adds diagnostic packet traces without changing that extraction contract.

Documentation

• Guild Wars DAT and snapshot format
• Archive decompression and texture payloads
• Skill extraction
• Item identity, localization, and icons
• NPC identity, localization, and vendor services
• Quest packets, localization, dialogue roles, and rewards
• Investigation journal

Legal and data ownership

TyriaExtractor is an independent, unofficial project. It is not affiliated with, endorsed by, or sponsored by ArenaNet or NCSOFT. Guild Wars, its trademarks, game files, text, artwork, and other game content remain the property of their respective owners.

The public source tree is intended to contain only the extractor source plus small synthetic fixtures used to document and validate decoder behavior. Gw.dat, Gw.snapshot, client executables, complete runtime captures, extracted catalogs, and extracted image collections must remain local and must not be committed or redistributed with the project. Users must provide their own official local client files; captures/ and all generated output/ artifacts are ignored by Git.

The MIT License applies only to original TyriaExtractor software and project material. It does not grant rights to Guild Wars data or assets extracted with the tools. Runtime instrumentation operates on the user's local client process; users are responsible for complying with applicable terms and laws.

References

The investigation and tooling were informed by these community projects:

• GWToolbox++ — Guild Wars client structures, GWCA interfaces, packet behavior, and mature runtime tooling.
• Py4GW — Python client integration and packet-sniffing patterns used during early capture experiments.

They are used as conceptual references and cross-checks. TyriaExtractor does not treat their legacy implementations as authoritative format definitions; exported game data still comes from the user's official local client.

License

Original TyriaExtractor software is licensed under the MIT License.