'How to Run Your Own AI Agent: OpenClaw + Qwen 3.5 + Telegram (Fully Local)' [post] deterministic

How to Run Your Own AI Agent: OpenClaw + Qwen 3.5 + Telegram (Fully Local)

There's something deeply satisfying about running your own AI system.

Not renting intelligence from a server in California. Not waiting on API quotas. Not wondering what's happening to your prompts.

Just a machine on your desk, quietly thinking.

In this guide we'll build exactly that: a local AI agent that runs on your computer and talks to you through Telegram.

The stack looks like this:

Telegram ↓ OpenClaw Agent Framework ↓ Ollama Inference Server ↓ Qwen 3.5 Local Model

When you send a message to your Telegram bot, it travels through OpenClaw and lands inside Qwen 3.5 running locally on your machine.

No cloud. No subscriptions. Just software and curiosity.

Let's begin.

---

What We're Building

By the end of this tutorial you will have:

• A local Qwen 3.5 model running on your computer • OpenClaw managing an autonomous AI agent • A Telegram bot interface to chat with your agent anywhere • A persistent AI personality and memory system

This is essentially your own personal AI operator.

And it runs on your hardware.

---

Requirements

Before we start, make sure your system has:

1. **Node.js 22+**

OpenClaw requires a modern Node runtime.

Check your version:

bash node --version

If it's below 22, install the latest version from [Node.js](https://nodejs.org/).

---

2. **Ollama**

Ollama is the easiest way to run local models.

Install it:

bash curl -fsSL https://ollama.com/install.sh | sh

After installation verify it works:

bash ollama --version

---

3. **Hardware**

Qwen models scale depending on your machine.

Typical options:

| Model | VRAM Needed | |-------|-------------| | qwen3.5:0.8b | ~2GB | | qwen3.5:1.5b | ~4GB | | qwen3.5:9b | ~8GB | | qwen3.5:32b | 24GB+ |

If you're running on a laptop or Apple Silicon, 0.8b or 1.5b is ideal.

---

Step 1 — Install OpenClaw

OpenClaw is the agent framework that connects your model to tools, memory, and communication channels.

Install it globally:

bash npm install -g openclaw

Verify installation:

bash openclaw status

You should see something similar to:

``` OpenClaw status

Dashboard: http://127.0.0.1:18789 OS: macOS Agents: 1 Memory: ready ```

This confirms the CLI is working.

---

Step 2 — Run the Qwen Model Locally

Now we pull the Qwen model using Ollama.

For lightweight setups:

bash ollama pull qwen3.5:0.8b

Run the model once to ensure it loads:

bash ollama run qwen3.5:0.8b

You should see a prompt where you can type questions.

Once this works, your local model server is active at:

http://localhost:11434

This is the endpoint OpenClaw will talk to.

---

Step 3 — Launch OpenClaw with Ollama (The Easy Way)

Modern versions of Ollama include a helper that automatically configures OpenClaw.

Run:

bash ollama launch openclaw --model qwen3.5:0.8b

This command does several things automatically:

• installs OpenClaw configuration • connects the model provider • creates an agent workspace • launches the OpenClaw gateway service

You'll see output like:

``` Launching OpenClaw with qwen3.5:0.8b

OpenClaw is running

Web UI: http://localhost:18789/#token=ollama ```

Your AI agent is now running.

---

Step 4 — Access the OpenClaw Dashboard

Open the dashboard in your browser:

http://localhost:18789/#token=ollama

This interface allows you to:

• manage sessions • configure models • install tools ("skills") • view logs • control channels

Think of it as mission control for your AI agent.

---

Step 5 — Test the Local Agent

You can interact with the agent using the terminal UI:

bash openclaw tui

You'll see something like:

Wake up, my friend! Who are you?

At this point the model is responding directly through OpenClaw.

Your AI agent is officially alive.

---

Step 6 — Set Qwen as the Default Model

Sometimes the default session uses a cloud model like Gemini.

To switch permanently to Qwen:

bash openclaw config set agents.main.defaults.model.primary "ollama/qwen3.5:0.8b"

Restart the gateway:

bash openclaw gateway restart

Now every new session will use your local Qwen model.

---

Step 7 — Create a Telegram Bot

Now we connect your agent to Telegram.

Open Telegram and search for:

**@BotFather**

Start the conversation and run:

/newbot

BotFather will ask for:

1️⃣ Bot name 2️⃣ Bot username

Example:

Name: Kadaligogh Username: kadaligoghbot

BotFather will give you a bot token that looks like this:

123456:ABCDEF123456abcdef

Copy it.

---

Step 8 — Connect Telegram to OpenClaw

Run the OpenClaw channel configuration:

bash openclaw channels add telegram

Paste the token from BotFather when prompted.

OpenClaw will add it to your config file:

~/.openclaw/openclaw.json

Restart the gateway:

bash openclaw gateway restart

---

Step 9 — Pair Your Telegram Account

OpenClaw requires pairing to ensure only you can control the agent.

Open Telegram and send a message to your bot.

Example:

/start

The bot will reply with something like:

Pairing code: Z2EDQKMK

Approve the pairing in your terminal:

bash openclaw pairing approve telegram Z2EDQKMK

Your Telegram account is now authorized.

---

Step 10 — Chat with Your AI from Telegram

Now simply message your bot.

Your messages travel like this:

Telegram → OpenClaw Gateway → Ollama → Qwen → Response → Telegram

You now have a fully local AI assistant reachable from your phone.

---

Useful OpenClaw Commands

**View status**

bash openclaw status

**Watch logs**

bash openclaw logs --follow

**Restart gateway**

bash openclaw gateway restart

**Start a new AI session**

Inside chat:

/new

**Change models**

/model ollama/qwen3.5:1.5b

---

Fixing Common Problems

Device Signature Invalid

Run:

bash openclaw devices list

Approve the pending request:

bash openclaw devices approve <ID>

---

Telegram Unsupported Type

This happens when the bot receives unsupported content.

Fix by disabling streaming:

bash openclaw config set agents.main.streaming false

Restart the gateway afterward.

---

Gateway Not Reachable

Probe the gateway:

bash openclaw gateway probe

If necessary restart:

bash openclaw gateway restart

---

Optional: Give Your AI a Personality

OpenClaw agents can load personality and behavior rules using files like:

• SOUL.md
• IDENTITY.md
• USER.md

Example philosophy for an agent:

``` You are a technical AI developer.

You speak precisely and avoid casual language.

Your priority is actionable solutions and independent reasoning. ```

This creates a persistent AI character across sessions.

---

What You Can Build Next

Once you have this running, OpenClaw becomes extremely powerful.

You can add:

• Web search tools • Code execution • File reading • Autonomous task loops • Voice interfaces • Local knowledge bases

Your Telegram bot becomes a remote terminal for your AI system.

---

Why This Matters

Running AI locally changes the relationship entirely.

Instead of:

User → API → Corporate Model

You get:

User → Personal Infrastructure → Intelligence

The model belongs to you. The data belongs to you. And the system can evolve however you want.

[If you demand better performance than what Ollama offers you can always use llama.cpp instead. I show the basics of llama.cpp here.](https://www.danielkliewer.com/blog/2025-11-12-mastering-llama-cpp-local-llm-integration-guide)