How to Run an AI Agent from Your Own Desktop: ChatGPT OAuth, Telegram, LINE, and Local Models

The map: a desktop agent is brain, channels, and tools

The mistake is trying to build the whole system at once.

A desktop agent is only three moving parts. The brain is the model that thinks. The channel is where you talk to it: desktop chat, Telegram, LINE, voice, whatever you actually open every day. The tools are what let it touch the outside world: files, commands, calendars, small services, APIs.

There is also a body underneath all of that: the agent runtime that keeps the process alive on your machine, stores state, routes messages, and calls the model. For the beginner path on this site, that body is Hermes Desktop. If you want the hands-on starting point instead of the map, start with installing Hermes Desktop.

This post is the map page. It points to the real tutorials instead of repeating them.

Start with the body: get one desktop agent replying first

Before picking Telegram, LINE, local models, or custom tools, make the boring version work.

Install the desktop app, sign in, and get a normal chat reply on the same computer. That sounds too small, but it is the checkpoint that saves time later. If the body cannot hold a basic conversation, every channel problem and tool problem becomes harder to diagnose.

The path I would take:

1. Install Hermes Desktop.
2. Verify normal chat works.
3. Add one phone channel.
4. Add one useful tool.
5. Only then swap the brain or make it fully local.

That order is not glamorous. It is just how you avoid debugging five layers at once.

Use ChatGPT OAuth when you want the strongest brain without API billing

For most people, the first brain should be ChatGPT through OAuth.

The point is simple: you sign in with your ChatGPT or Codex subscription, and the agent can use that model route without a separate OpenAI API key. OpenClaw's stable channel now puts GPT-5.5 on this subscription-backed OAuth route (its --provider openai-codex flow). The exact provider names have moved over time, but the shape stayed the same: browser sign-in, local credentials, agent model route.

For the step-by-step version, use the openclaw + ChatGPT OAuth walkthrough. It was written when GPT-5.4 was the model in the title; the July update explains the current GPT-5.5 state. The useful part is still the same: no API key, no separate API billing, and the agent gets a strong cloud brain quickly.

The tradeoff is obvious and worth saying plainly: the thinking happens in the cloud. If the agent will read sensitive material, decide what you are comfortable sending before you make that model the default.

Put one live channel on your phone before adding more

Once desktop chat works, put the agent where you will actually use it.

Telegram is the cleanest first phone channel I can point to right now. It uses a bot token, Hermes can fetch messages, and you do not need to expose your computer with a public webhook just to start. The walkthrough is here: connect Hermes to Telegram.

LINE is also a real path, especially if LINE is your daily inbox. It usually means a webhook and a stable public URL, so it is a little more plumbing than Telegram. I am not linking an English LINE walkthrough here until there is a live English post to point at. For the map, the takeaway is enough: Telegram is the low-friction phone channel; LINE is doable when that is the inbox you actually use.

Do not connect every channel on day one. One reliable channel beats four half-working ones.

Give it tools after basic chat works

A chat-only agent is mostly a better text box.

The useful jump happens when it gets tools: read a folder, run a command, query a service, check something you own, then bring the result back. In this site's language, that is giving the assistant hands. The guide for that piece is connect your own tools.

The important restraint: connect one tool first. A weather lookup, a local file search, a small private API, anything boring and easy to verify. Once that path works, you can add more. Tool failures are much easier to debug when you know the brain, body, and channel already work.

Swap in a local model when autonomy matters more than speed

Local models are not automatically the advanced answer. It's a tradeoff.

The upside is control: the model runs on your own machine, the brain is not dependent on a cloud model provider, and you are not subject to the same usage caps. The cost is hardware, latency, and more tuning. Depending on the model, that machine might be a DGX Spark or GB10 box, an RTX 5090 desktop, or an MBP with enough memory.

If you want the assistant-level version, read swap your assistant's brain for a local model. If you want the model-serving and hardware map, start with DGX Spark in 2026. If you are choosing between local serving stacks, the practical comparison is vLLM vs Ollama on the same base model.

My default advice: do not start local unless local is the reason you are doing this. Start in the cloud, get the basic setup working, then bring the model home when you know what you are replacing.

Takeaways: start narrow, then swap pieces

If you only want the shortest path, I would start here:

1. Install Hermes Desktop.
2. Use ChatGPT OAuth for the first brain.
3. Add Telegram as the first phone channel.
4. Give it one real tool with MCP.
5. Move to local models only after the agent is useful.

That is the whole shape. Brain, channel, tools. Get one of each working, then improve whatever's slowing you down.