Skip to content

tars ai.stub

Nicolas Cravino edited this page Jul 10, 2026 · 4 revisions

id: tars-ai name: TARS-AI repo_path: ~/Documents/sw30labs/repos/tars-ai remote_url: https://github.com/sw30labs/tars-ai.git primary_language: Python framework: LangGraph + OMLX (OpenAI SDK) category: local-inference-mlx stacks: [mlx, langgraph, agentic] ingested: 2026-04-11 last_commit_date: 2026-07-06

TARS-AI

Purpose

LangGraph-powered conversational AI embodying TARS from Interstellar. Fully local, no cloud APIs — text in, TARS voice out. As of the July 2026 refactor, both the LLM and the TTS run on a local OMLX server (OpenAI-compatible MLX inference); the client holds no models and downloads nothing.

Architecture

LangGraph agent graph with three nodes — Personality Injector → LLM Node → Voice Output — where both the LLM and Voice nodes are OpenAI-SDK clients pointed at a local OMLX server (http://127.0.0.1:8000):

  • Personality Injector: configurable settings (humor, honesty, discretion, trust), adjustable at runtime ("Set humor to 50")
  • LLM Node: OMLX /v1/chat/completions, default model MiniMax-M2.7-ultra-uncensored-heretic-oQ4-MLX
  • Voice Output: OMLX /v1/audio/speech using Qwen3-TTS Voice Design — the voice is generated from a text description (sent as the OpenAI instructions field), not a fixed preset

This replaces the earlier in-process design (mlx-lm Llama 3.1 8B 4-bit + mlx-audio Kokoro-82M); models now live on the OMLX box and load on demand. Text-only mode (--no-voice) runs anywhere.

Key Technologies

  • Framework: LangGraph, OMLX (OpenAI-compatible MLX server)
  • LLM: served by OMLX (MiniMax-M2.7 heretic, MLX)
  • TTS: Qwen3-TTS Voice Design via OMLX /v1/audio/speech
  • Client SDK: OpenAI Python SDK

Dependencies

openai, langchain-core, langgraph, sounddevice, soundfile, numpy, pydantic, rich, pyyaml (LLM + TTS offloaded to the OMLX server; no mlx-lm/mlx-audio in the client)

Tags

python, langgraph, mlx, apple-silicon, local-inference, tts, llm

Cross-references

Clone this wiki locally