Self-Hosted AI Agent Orchestration: A Full Tour

What This Video Covers

A 24-minute walkthrough of TheYgent on one Mac, from git clone to a published agent. It opens with what the platform is — a no-code, local-first studio with built-in llama.cpp, MLX, and vLLM inference, MCP tools, retrieval sources, run traces, durable runs, and Docker Compose or Kubernetes deployment — and a quick look at the repository, the user docs, and the architecture: an interface, a control plane, an inference plane, a durable worker, and Postgres.

Then it installs the stack (cp .env.example .env, make engines, make up), creates the admin account in the first-run wizard, installs a model from the in-app Hugging Face catalog — each variant shows whether it fits the machine — and tries it on the Bench, where parameters can be tuned and saved as presets.

On the canvas it builds a first agent: an input, a local LLM, and an output, tested in the side panel with a trace of every run. Then it adds an MCP tool (a local Docker server) that the model decides to call on its own, and publishes the agent as a versioned API. The tour ends with runs and chat, MCP servers (hubs, stdio, remote, and OpenAPI or GraphQL specs turned into MCP), retrieval sources, the settings — samples, users, sign-in, inference, telemetry, credentials — and import and export.

Try It Yourself

• Installation
• Your first agent
• Installing local models
• MCP servers