mybrain.tv

What It Is

mybrain.tv is an experimental AI coworker project built around the idea of the “Centaur Way,” with its core product concept called buddy. Rather than positioning AI as a passive Q&A tool, it tries to make AI behave more like a team member, with shared goals, defined responsibilities, progress reviews, and the ability to re-plan. The site also publishes its design discussions, failures, and human feedback process in a serialized format, making the project itself a validation experiment in human-AI co-creation.

Core Capabilities and Architecture

The article discloses a fairly concrete technical stack: local inference uses Qwen3.5-35B MLX, speech synthesis / voice clone uses Qwen3-TTS MLX, speech recognition is handled by faster-whisper, and nomic-embed provides embedding capabilities via Docker. The system includes an orchestrator / IPC layer, sub-agents such as coder / researcher / writer, and a real-time voice interface based on WebRTC / OpenVidu. The memory layer uses an MD-first + Git approach, combined with an embedding index for semantic retrieval. In terms of product mechanics, buddy commits to tasks around OKRs, periodically checks progress via Heartbeat, and re-plans on its own when needed.

Pricing and Availability

The site does not disclose any SaaS subscription, one-time license, open-source license, or public trial entry point. Its value proposition is “自分の手元で動かすこと” — running it in your own environment — with an emphasis on Apple Silicon local deployment and zero cloud usage fees. The example development environment in the article is an Apple M5 Max with 128GB of unified memory, which means inference costs may be controllable, but the hardware barrier is not low.

Pros and Cons

The main advantage is a very clear privacy boundary: conversations are not sent to third parties, making it suitable for individual R&D or experimental scenarios with high data confidentiality requirements. The architecture covers LLM, TTS, STT, memory, agent orchestration, and real-time voice, giving it a complete and transparent direction. The limitations are also obvious: at present, it feels more like a proof of concept being built in public, with no installation documentation, product versions, performance benchmarks, stability data, enterprise permissions, or support information. Output quality still depends on continuous human feedback and process design.

Who It’s For and Access from China

It is better suited to AI Agent developers, local model enthusiasts, small teams studying human-AI collaboration workflows, and people who want to build privacy-first AI assistants. It is not ideal for mainstream users expecting an out-of-the-box SaaS product. Access from mainland China, payment methods, and network availability are not explained in the article, so they should be considered unknown. Comparable or reference directions include Ollama, LM Studio, AnythingLLM, Dify, LangGraph, CrewAI, AutoGen, Open Interpreter, and others.