TedBot – OpenClaw Lite for Local AI Agents

Remote bot via telegram

Local only interactive platform for displays, iPads, smartphone

---

TL;DR 🚀

TedBot is a personal AI agent for <$100 setup that:

• Turns natural language → calendar events, notes, and reminders
• Runs locally (CPU only) with minimal overhead
• Connects to Telegram for lightweight automation

No cloud. No vendor lock-in. No overhead.

---

Think of it as :

"If you wanted a fully private assistant to manage your reminders and notes, this is the solution for you"

---

Example Use Cases

These are intentionally simple but representative:

Calendar intent extraction

“Remind me tomorrow at 9am” → structured time + action

Note summarization

"For Project X, make sure I read up on tokenisation methods" →Turning unstructured text into concise notes

Action extraction

"Add it to work category" Identifying tasks or follow-ups from text

🚩 Architecture

        +----------------+
        | Telegram / CLI |
        +--------+-------+
                 |
                 v
          +------+------+
          |  TedBot AI  |
          |  (OpenClaw  |
          |   Lite)     |
          +------+------+
                 |
         +-------+-------+
         | Calendar/Notes|
         +---------------+

---

Quick Start

# 1. Clone
git clone <repo>

# 2. Start system
docker compose -f docker-compose.tedbot.yml up -d

# 3. Run model
ollama run gemma3:1b

---

🔧 TedBot Troubleshooting Guide

This guide captures the real-world debugging commands, memory checks, and container orchestration tricks from the TedBot setup.

---

1. Ollama Model Management

Issue	Commands / Checks	Notes / Outcome
Check running models	ollama ps	Lists all currently running Ollama models with memory usage.
Stop a misbehaving model	ollama stop <model>	Stops the model gracefully; check ollama ps afterwards.
Remove unwanted models	ollama rm <model>	Cleans up disk usage; useful when upgrading models.
Run a model	ollama run <model>	Start a model for inference. Combine with curl for API testing.
Pull / update a model	ollama pull <model>	Always check free memory before pulling large models.

💡 Pro tip: Use ollama list to see all downloaded models, ollama show <model> for metadata.

---

2. Docker & Docker Compose

Issue	Commands / Checks	Notes / Outcome
Start services	docker compose -f docker-compose.tedbot.yml up -d	Runs TedBot in detached mode.
Restart specific container	docker restart <container>	Useful if UI crashes but backend is fine.
Check logs	docker logs -f <container>	Continuous output; essential for debugging crashes or misfires.
Down & rebuild	docker compose -f docker-compose.tedbot.yml down && docker compose -f docker-compose.tedbot.yml up -d	Resets the container environment; solves many networking or volume issues.
Verify running containers	docker ps -a	Ensures expected containers exist and are in the correct state.

💡 Tip: Always docker logs before restarting — avoids losing error context.

---

3. System Resources & Memory

Issue	Commands / Checks	Notes / Outcome
Check RAM usage	free -h	Useful before starting large models; ensures you won’t hit OOM.
Monitor CPU usage	top / `watch -n 1 'ps aux --sort=-%cpu	head -15'`
Number of cores	nproc	Helps configure threading for Ollama models.
CPU model	`cat /proc/cpuinfo	grep "model name"
Power info	sudo dmidecode -t 39 / `sudo dmidecode -t 39	grep -i watt`

💡 Pro tip: Monitor memory and CPU while running gemma3:1b or qwen2.5:3b; Ollama is CPU-bound on small servers.

---

4. Ollama API Testing

Issue	Commands / Checks	Notes / Outcome
Test API locally	curl -s -o /dev/null -w "%{time_total}" http://localhost:11434/api/generate -d '{"model":"qwen2.5:3b","prompt":"hi","stream":false,"options":{"num_predict":5}}'	Confirms that the model responds correctly over HTTP.
Verify environment	systemctl show ollama --property=Environment	Ensures systemd service uses correct host/port and keep-alive settings.

---

5. Systemd Service Tweaks (Optional)

Issue	Commands / Checks	Notes / Outcome
Override Ollama environment	sudo tee /etc/systemd/system/ollama.service.d/override.conf	Set host, port, or keep-alive options.
Reload & restart service	sudo systemctl daemon-reload && sudo systemctl restart ollama	Applies changes without rebooting.

💡 Pro tip: Use environment overrides to expose Ollama for remote API calls safely.

---

6. Common Pitfalls

• Memory Errors: Always free -h before running 3B+ models. If OOM, stop unused models (ollama stop <model>).
• SSH Copy Confusion: Don’t run scp from the remote server; always run from local host.
• Container Networking: If API calls fail, check docker ps and container port bindings.
• Zombie Containers: Periodically run docker ps -a and docker rm old containers.
• Logs Disappear: Always run docker logs -f tedbot before restarting the container.

---

At a Glance

TedBot is not a single script — it is a composed system:

UI (Dockerized)

• Handles user interaction and input

Backend (TedBot service)

• Responsible for:
  • request normalization
  • prompt construction
  • response parsing (e.g., extracting structured intent)

Ollama (Local model runtime)

• Runs models like gemma, qwen, etc. locally

System Layer

• Provides:
  • process visibility
  • logs
  • latency measurement
  • resource monitoring

This separation is intentional — it mirrors how production AI systems are structured.

---

What Problem This Solves

Most LLM applications today:

- depend heavily on external APIs or require you to buy a Mac mini or equivalent device which costs over $100

- hide core behavior behind abstractions

- are difficult to debug or reason about

- do not run reliably in constrained environments

TedBot addresses this by:

- running fully local (no cloud dependency)

- exposing the entire request → inference → response pipeline

- making model behavior observable and debuggable

- operating within tight CPU and memory constraints-

The result is a system that prioritizes control, transparency, and reliability over raw capability.

💡 What Makes This Different

Most LLM projects:

• API wrappers
• Black-box demos
• Hardcoded logic

This one:

• Runs fully locally
• Shows the entire stack
• Lets you debug the LLM layer
• Built for endurance, not demos

---

🚡 Performance Reality

• CPU bound by design
• Model size tradeoffs
• Smaller models = faster iteration
• Observability > raw speed

---

💡 Real Use Cases (this is where it shines)

• "Remind me tomorrow at 9 AM" → calendar intent
• "Summarize my notes" – KNOWLEDGE layer
• "Extract action items" – agent behavior and retrieval

---