NPCTalk

Built first for GTA V: walk up to anyone in Los Santos and just talk. Lamar roasts your hoodie, a cop tells you to file a civil claim and flips you off, Michael defends his robe like it's a business suit. Every line is written live in character and spoken in that character's cloned voice by Fish Audio Drama 3 (or S2 Pro), self-hosted on a serverless GPU. Most of the code doesn't know it's in GTA, so the same pieces work in other games: the skills say what a game has to provide, and how to build the rest.

Install

Pick your agent. Each gets the same skills (Agent Skills format), the fal MCP server and the npctalk CLI.

Inside a clone, .agents/skills and .claude/skills are copies of skills/, and AGENTS.md holds the instructions (CLAUDE.md and GEMINI.md point to it).

The npctalk CLI. Plugin installs and clones put it on PATH (bin/). Anywhere else: uv tool install git+https://github.com/christos-antono/NPCTalk.

Accounts and keys (environment variables, never files):

You also need Git, Python 3.10+, uv and ffmpeg. The GTA V example needs Windows, the .NET SDK (8+) to build it, and the game.

Try it

Make the villagers in my game answer when I talk to them, in their own voices.

Set up a Fish Audio Drama 3 voice server on Modal and clone a voice from narrator.wav.

In GTA V story mode, let me talk to anyone near me with my mic.

My NPC's cloned voice sounds thin. Is it the reference clip or the model? Measure it.

Record five 8-second clips of NPCs arguing that look like normal gameplay.

The agent starts with the make-npcs-talk skill and runs the same loop every time:

1. recon the game: version, how code runs in it, single-player or not;
2. pick the engines (voice, LLM, speech-to-text, actions);
3. stand up the voice server first, outside the game, and test it with npctalk say;
4. make reference voices from recordings you have the rights to;
5. one NPC answering typed text in the game: the four hooks, nothing else;
6. voice input (push-to-talk, then the open mic);
7. characters: personas, their own lines, scene facts;
8. actions, and doing what they're told;
9. verify in the real game, with every step's timing logged;
10. show it, and measure the voices before claiming anything;
11. write a field note for the next agent.

What's inside

Skills (skills/, Agent Skills format)

The npctalk CLI (Python). Every command has --help with examples.

The voice server (server/): drama3.py (Fish Audio Drama 3, the default), s2pro.py (S2 Pro, with fine-tuned voices), keepwarm.py, fal/ (S2 Pro on fal over a realtime WebSocket), finetune/ (dataset, LoRA, and the scorer behind every number here).

The GTA V example (examples/gta-v): the full mod (open mic, any NPC, 19 characters, homes, actions, dev mode, a director for recording clips), its tools, and a field note.

Field notes (knowledge/): one per game where NPCs were made to talk. Add yours.

How good are the voices?

Measured on one character (Lamar, ten test lines): voice match is speaker similarity to 60 real clips that were never used as references. Details and the method: skills/voice-likeness-eval.

Without training, Drama 3 and S2 Pro tie on likeness; Drama 3 acts [crying] and [laughing] better and starts talking faster. One early result here didn't survive a second run: how that was caught.

Rules it follows

• Single-player games, or servers you run. It never injects into online games with anti-cheat (GTA V: never launch GTA Online with a script hook installed) and never bypasses anti-cheat, DRM or ownership checks.
• Voices you have the rights to. It never commits or publishes voice clip