← Back to Blog
Tool - 2026

AI Exodus Companion - A Voice for Your Game

📅 November 22, 2026 ⏱️ 6 min read
I play a lot of single-character games, and the loneliest moment is the quiet one: a long grind, nobody to say "nice kill" or "that quest was nonsense". I wanted a companion that sits beside the game, watches what happens and talks about it out loud, in first person, as the character I am playing. AI Exodus Companion is that: a Windows app for World of Warcraft, AION 2 and Fortnite. Here is how it works, and the problems that shaped it.
🎮 See AI Exodus Companion at ai-exodus.com →

What I Set Out To Build

A character that reacts, not a chatbot you have to talk to. It should notice a kill, a level-up or a quest turn-in, say something short in the right voice, remember what happened last session and, if you press the microphone, answer you. Three rules were non-negotiable from day one: it never sends input to the game, it never reads the game's memory, and it never asks for the game login. Everything it knows comes from what the screen shows, what the game writes to disk, or, in one case, a passive copy of the network traffic (more on that below).

The Shape Of The App

  • One Electron app, one pack per game. The core handles capture, speech, memory and the overlay. Each game (WoW, AION 2, Fortnite) is a pack that knows how to read that game.
  • Lines come from Gemini through my own server. The model key was removed from the player's PC, so the account server holds it and can meter usage.
  • The voice runs locally. Text-to-speech uses the Chatterbox Turbo model in ONNX, bundled in the installer, so no Python install is needed. Speech recognition for talk-back is a local Whisper model.
  • Memory is per character. The app keeps what happened and writes a nightly "session book" you can read afterwards.

Getting Events Out Of A Game That Does Not Want To Tell You

Each game exposes a different amount, so the reading method differs per game:

  • World of Warcraft: a small in-game add-on paints a strip of coloured blocks that encodes game state, and the app decodes that strip from its own screen capture. The add-on never sends input, never reads memory and never touches the network. The app also reads the game's chat log.
  • Fortnite: a log file plus HUD pixels. Eliminations, health and storm are not in the logs, so they come from the screen.
  • AION 2: the game has no log and no add-on, and the screen only shows health bars and a feed. For combat, a separate utility process takes a passive copy of the game's network traffic through Npcap, decodes hits, heals and deaths, groups them into fights and hands the app only finished facts. It sends nothing and never opens the game process, but it is reading traffic rather than the screen, and I think a player should know that. It can be switched off in Settings.
  • Text on screen: reading uses the Windows built-in OCR, which takes 2 to 18 ms per crop. My first OCR engine took 100 to 400 ms, which is too slow to run often.

The Problems That Took Longest

  • The pixel strip breaks when players change their graphics options. WoW's Brightness and Contrast apply first and Gamma after, so a colour I encoded as one value arrived as another. I wrote a solver that works out the brightness, contrast and gamma curve from known reference blocks inside the strip itself, then decodes. The rule is that a frame that does not match is refused, never guessed. Tested offline on 342 real frames the old decoder had refused: 308 decoded, none wrong.
  • The model was slow because "thinking" quietly switched on. I asked for the minimum thinking depth on gemini-3.8-flash, which the API refuses, and the server's retry dropped the setting, so lines ran at the default depth: about 1,150 thinking tokens and a median of 5.8 s, against 1.4 s when pinned to low (146 calls). Pinned to low, and every remark now asks for it.
  • A voice that arrives late is worse than silence. Early speech synthesis ran at about a quarter of real time. The fixes were pre-rendered reaction clips for the common moments (kill, hurt, aggro), cached per voice, then moving the model on-device. On a 2017 Ryzen 7 1700X the app starts speaking about 75 ms after it decides what to say and transcribes 10 seconds of speech in about 1.6 seconds.
  • Repetition and cost. A companion that repeats itself is annoying, so lines are checked against recent ones, sounds have a minimum gap, and quest-giver lines are stored with their audio, so a repeated quest makes no model call at all.
  • Counting what a player used. Usage is billed in "lines". The count lives on the server in a place the player's own credentials cannot write to, with an hourly burst limit and daily limits so a bug or a script cannot run up a bill.

Voice, Language And Memory

Voices can be cloned from a short read-aloud sample of about 25 seconds, using a speaker encoder that runs locally and is cached per voice. The multilingual model covers many languages, and the idle lines have French versions written natively rather than machine-translated. Voice matters more than I expected: the same line in a flat voice reads as spam, and in a voice you picked it reads as a character.

Pricing

The Companion is billed by lines written, not by hours played: Free is €0 for 150 lines a month, Companion €12 for 1,500 and Plus €20 for 3,500, per the pricing page.

What Is Not Finished

  • Talk-back recognition is English-only today: the bundled model is whisper-base.en. A multilingual speech model is supported by the code but is not shipped yet, so for now it has to be added by hand, and I have not verified French talk-back end to end.
  • The installer is large (over 1 GB) because the voice and recognition models ship inside it, and the app uses around 2 GB of memory.
  • The installer is not code-signed (executable signing is switched off in the build), so Windows may show a warning when you run it.
  • Each remark prompt contains a randomised line (a pet-name and address roll) early in the text, about 5,000 characters in. That stops any provider-side prompt cache from matching the static part before it. My own audit found this on 2026-10-26 and I have not changed it since.
  • Each new game needs its own reading method, so adding a game is real work, not a setting.

Want a companion for your own game, or help with overlays, OCR or on-device voice? Use the quote request form on the home page, or ask the assistant there.

🎮 See AI Exodus Companion at ai-exodus.com →