Building ollacode: A Local AI Coding Assistant with Telegram Integration (Day 1)

What is ollacode?

ollacode is a lightweight, local AI coding assistant that runs entirely on your machine. Think of it as a personal “Claude Code” or “Aider” — but using Ollama with the qwen3-coder:30b model, and with the bonus of Telegram integration for remote access.

No cloud APIs. No subscription fees. Just your local GPU doing the work.

Why Build This?

I’ve been impressed by tools like Claude Code and Aider, but I wanted something:

1. Fully local — My code stays on my machine
2. Telegram-enabled — Ask coding questions from my phone while away from my desk
3. Lightweight — No heavy IDE plugins, just a terminal command
4. Customizable — My own tools, my own rules

Architecture

Here’s how ollacode is structured:

graph TD
    CLI["🖥️ CLI - Rich<br/>Streaming + Approval UI"] --> Engine
    TG["📱 Telegram Bot<br/>Per-user Sessions"] --> Engine
    Engine["⚙️ Conversation Engine<br/>History | Tool Orchestration<br/>Agentic Loop | Project Memory"]
    Engine --> Ollama["🔗 Ollama Client<br/>httpx async"]
    Engine --> Tools["🛠️ Tool System<br/>7 tools"]
    Engine --> Prompts["📋 System Prompt<br/>+ OLLACODE.md Memory"]
    Ollama --> Server["🧠 Ollama Server<br/>localhost:11434<br/>qwen3-coder:30b"]

    style CLI fill:#4a9eff,stroke:#2d7cd4,color:#fff
    style TG fill:#0088cc,stroke:#006699,color:#fff
    style Engine fill:#9b59b6,stroke:#7d3c98,color:#fff
    style Ollama fill:#e67e22,stroke:#d35400,color:#fff
    style Tools fill:#27ae60,stroke:#1e8449,color:#fff
    style Prompts fill:#f39c12,stroke:#d68910,color:#fff
    style Server fill:#2c3e50,stroke:#1a252f,color:#fff

The key design decision was separating the Conversation Engine from the interfaces. Both CLI and Telegram share the same engine, so all the smart logic (tool calling, agentic loops, project memory) works identically regardless of how you interact with it.

What I Built on Day 1

Core Foundation

I built the entire working system in a single session:

7 Python modules, ~1500 lines of code:

Module	Purpose
config.py	Settings management via .env files
ollama_client.py	Async HTTP client for Ollama’s /api/chat API
engine.py	Conversation engine with agentic loop
tools.py	7 coding tools (file ops, search, command execution)
prompts.py	System prompt + project memory loader
main.py	CLI interface with Rich terminal UI
telegram_bot.py	Telegram bot with per-user sessions

The Tool System

This is where the magic happens. The AI doesn’t just chat — it can act:

Tool	What It Does
read_file	Read files with line numbers
write_file	Create new files
edit_file	Diff-based search/replace editing
list_directory	Browse directory contents
search_files	Find files by glob pattern
grep_search	Search inside file contents
run_command	Execute shell commands

The edit_file tool is particularly important — inspired by Aider’s approach, it uses search/replace blocks instead of rewriting entire files. This means:

• Token efficiency: Only the changed portion is sent
• Safety: No risk of accidentally wiping a file
• Diff preview: You see exactly what will change before it happens

Agentic Loop

The conversation engine runs an agentic loop — up to 10 iterations of:

flowchart LR
    A["👤 User Request"] --> B["🤖 AI Response"]
    B --> C{"Tool Calls?"}
    C -- No --> D["✅ Return Response"]
    C -- Yes --> E["⚙️ Execute Tools"]
    E --> F["📊 Feed Results Back"]
    F --> G["🤖 AI Analyzes"]
    G --> H{"More Tools?"}
    H -- Yes --> E
    H -- No --> D

    style A fill:#3498db,stroke:#2980b9,color:#fff
    style B fill:#9b59b6,stroke:#7d3c98,color:#fff
    style D fill:#27ae60,stroke:#1e8449,color:#fff
    style E fill:#e67e22,stroke:#d35400,color:#fff
    style F fill:#f39c12,stroke:#d68910,color:#fff
    style G fill:#9b59b6,stroke:#7d3c98,color:#fff

If a tool returns an error, the engine automatically prompts the AI to analyze and fix the issue. This means the AI can:

1. Read a file
2. Edit it
3. Run tests
4. See test failures
5. Fix the code
6. Re-run tests
7. Report success

All in a single user request.

Approval System

Inspired by Cline’s human-in-the-loop approach, dangerous operations require user approval:

• CLI: Shows a diff preview + y/n/a prompt before file modifications or command execution
• Telegram: Auto-approves for convenience (since you’re already remote)
• --auto-approve flag: Skip all prompts when you trust the AI

Project Memory (OLLACODE.md)

Inspired by Claude Code’s CLAUDE.md, drop an OLLACODE.md file in your workspace root:

# Project Rules
- Python 3.12, type hints required
- Use pytest for testing
- Database: PostgreSQL with SQLAlchemy

This gets automatically injected into every conversation, so the AI always knows your project’s conventions.

Tech Stack

Layer	Technology
Language	Python 3.10+
LLM Backend	Ollama (local)
Model	qwen3-coder:30b
HTTP Client	httpx (async)
CLI UI	Rich + prompt-toolkit
Telegram	python-telegram-bot
Config	python-dotenv

Quick Demo

CLI Mode:

$ ollacode cli
   ____  _ _         _____          _
  / __ \| | |       / ____|        | |
 | |  | | | | __ _ | |     ___   __| | ___
 | |  | | | |/ _` || |    / _ \ / _` |/ _ \
 | |__| | | | (_| || |___| (_) | (_| |  __/
  \____/|_|_|\__,_| \_____\___/ \__,_|\___|

ollacode ❯ Fix the bug in main.py

Telegram Mode:

$ ollacode telegram
🤖 ollacode Telegram 봇을 시작합니다...
   모델: qwen3-coder:30b
   서버: http://localhost:11434

Then just message your bot from Telegram — ask it to review code, write functions, debug errors, all from your phone.

What’s Next (Day 2+)

Based on my research of Claude Code, Aider, and Cline, here’s what I’m planning:

• Git integration — Auto-commit with AI-generated messages
• Context management — Auto-summarize long conversations to save tokens
• Codebase awareness — Auto-detect project structure and frameworks
• Plan mode — Outline steps before execution
• Plugin system — User-defined custom tools

Get It

git clone https://github.com/rockyRunnr/ollacode.git
cd ollacode
python3 -m venv .venv && source .venv/bin/activate
pip install -e .
cp .env.example .env
ollacode cli

Requires Ollama with qwen3-coder:30b pulled.

---

This is Day 1 of the ollacode devlog. Follow along as I build a full-featured local AI coding assistant.